<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Samir Jha|Oracle Architect (OCI|OIC|SaaS Appl) on Medium]]></title>
        <description><![CDATA[Stories by Samir Jha|Oracle Architect (OCI|OIC|SaaS Appl) on Medium]]></description>
        <link>https://medium.com/@jhasamiree1?source=rss-8a00b14ded6e------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*HETEBau9nbzmhnbk.jpg</url>
            <title>Stories by Samir Jha|Oracle Architect (OCI|OIC|SaaS Appl) on Medium</title>
            <link>https://medium.com/@jhasamiree1?source=rss-8a00b14ded6e------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Fri, 20 Nov 2026 00:57:54 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@jhasamiree1/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Oracle Fusion — Data Conversion:Purchase Order]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/@jhasamiree1/oracle-fusion-purchase-order-data-conversion-22c0132cc552?source=rss-8a00b14ded6e------2"><img src="https://cdn-images-1.medium.com/max/1536/1*y1o2l9Sj7ahXpgPNzVkBAQ.png" width="1536"></a></p><p class="medium-feed-snippet">This blog is co-authored by Kshemesh Mehta</p><p class="medium-feed-link"><a href="https://medium.com/@jhasamiree1/oracle-fusion-purchase-order-data-conversion-22c0132cc552?source=rss-8a00b14ded6e------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://medium.com/@jhasamiree1/oracle-fusion-purchase-order-data-conversion-22c0132cc552?source=rss-8a00b14ded6e------2</link>
            <guid isPermaLink="false">https://medium.com/p/22c0132cc552</guid>
            <category><![CDATA[purchase-orders]]></category>
            <category><![CDATA[data-conversion]]></category>
            <dc:creator><![CDATA[Samir Jha|Oracle Architect (OCI|OIC|SaaS Appl)]]></dc:creator>
            <pubDate>Mon, 05 Oct 2026 11:25:01 GMT</pubDate>
            <atom:updated>2026-10-06T13:37:33.244Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[Hybrid Graph RAG: From a Simple PDF Chatbot to Intelligent Search]]></title>
            <link>https://medium.com/@jhasamiree1/hybrid-graph-rag-from-a-simple-pdf-chatbot-to-intelligent-search-da2c740bbbcb?source=rss-8a00b14ded6e------2</link>
            <guid isPermaLink="false">https://medium.com/p/da2c740bbbcb</guid>
            <dc:creator><![CDATA[Samir Jha|Oracle Architect (OCI|OIC|SaaS Appl)]]></dc:creator>
            <pubDate>Sun, 04 Oct 2026 10:49:33 GMT</pubDate>
            <atom:updated>2026-10-05T11:50:13.887Z</atom:updated>
            <content:encoded><![CDATA[<h3>Why I built this project</h3><p>I had been reading about RAG, embeddings, vector databases, grounding and AI agents, but I found that the concepts stayed abstract until I tried to build something myself.</p><p>Instead of building another generic chatbot, I chose an enterprise-style problem that was easy to understand: a supplier policy assistant. The idea was simple. A user uploads a supplier policy PDF, asks a question, and the application returns an answer grounded in the policy document.</p><p>I built the first version as a traditional Vector RAG application. Then I kept asking a natural question: what happens when the answer is not only about similar text, but about relationships between business concepts? That question led me to Knowledge Graphs and eventually to Hybrid Graph RAG.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/619/1*rIgj5Hgm49niVYZ_Ptw4_w.png" /></figure><h3>What is RAG?</h3><p>Before writing code, I reduced RAG to one simple idea: do not ask the language model to remember the policy. Retrieve the relevant policy text first, then give that text to the model as context.</p><p><strong>Definition of grounding</strong><br> Grounding means the answer is based on retrieved source context rather than the model inventing an answer from general knowledge.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/907/1*z8EBrEFb0XYpzwIKqMp-fg.png" /><figcaption><em>Figure 2. The final high-level architecture: document processing feeds both the vector store and knowledge graph, then hybrid retrieval supplies context to GPT-5-mini.</em></figcaption></figure><h3>Traditional Vector RAG came first</h3><p>The first architecture was intentionally small: PDF → text extraction → chunks → embeddings → FAISS → relevant chunks → LLM. This was enough to understand the core RAG loop.</p><p>User question<br> ↓<br> Create question embedding<br> ↓<br> Search FAISS<br> ↓<br> Retrieve relevant chunks<br> ↓<br> Send chunks + question to GPT<br> ↓<br> Grounded answer</p><p>This version taught me the most important RAG building blocks: embeddings, chunking, similarity search, context retrieval and grounding. It also introduced the idea of provenance — being able to say which document and page supported an answer.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*jlhksJV0tM4P9BQFREXgoA.png" /></figure><h3>Why move beyond Vector RAG?</h3><p><strong><em>Vector search </em></strong>is excellent at finding text that is semantically similar to a question. But enterprise questions often involve explicit relationships: <em>which department approves something, which supplier category a payment term applies to, or what action happens when an invoice is non-compliant.</em></p><p>That led me to a second mental model: the vector store answers “<strong><em>what text is similar?</em></strong>”, while the graph answers “what business concepts are connected?”</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*H73n3d4EMzDAmWP_qaLcWA.png" /><figcaption><em>Figure 3. Examples of the business relationships represented in the graph.</em></figcaption></figure><h3>The Hybrid Graph RAG idea</h3><p>The final application combines both approaches. Vector retrieval finds the supporting policy text. Graph retrieval finds structured relationships when those relationships are relevant to the question. The two evidence types are then combined before the LLM generates the answer.</p><p>Question<br> ├── Vector retrieval → relevant policy chunks<br> └── Graph retrieval → relevant business relationships<br> ↓<br> Hybrid context<br> ↓<br> GPT-5-mini<br> ↓<br> Answer + provenance</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ZVij9wsI32OYmFujoJX0rg.png" /></figure><h3>How I built it, step by step</h3><p>The biggest lesson for me was to build one responsibility at a time and test it before moving forward. Here is the actual sequence I followed.</p><p><strong>1. Created the project structure</strong> — created separate backend/ and frontend/ folders. The backend would own the RAG logic and the frontend would provide the React interface.</p><p><strong>2. Created the Python environment</strong> — created backend/.venv and installed FastAPI, Uvicorn, python-multipart, python-dotenv, openai, pypdf, faiss-cpu and networkx.</p><p><strong>3. Configured OpenAI-</strong>created backend/.env for the API key and tested text-embedding-3-small. The embedding response returned 1536 dimensions, which became the FAISS index dimension.</p><p><strong>4. Added PDF extraction</strong> — pdf_service.py extracts text page by page. Preserving the page number early made later provenance possible.</p><p><strong>5. Added chunking</strong> — chunk_service.py split page text into 1200-character chunks with 200-character overlap. We also fixed the end-of-text loop so the final chunk could not cause an infinite loop.</p><p><strong>6. Added embeddings</strong> — embedding_service.py became the single place responsible for calling OpenAI embeddings.</p><p><strong>7. Added persistent FAISS storage</strong> — vector_store.py stored vectors and metadata on disk so the vector index survives a backend restart.</p><p><strong>8. Added the knowledge graph</strong> — graph_service.py created a NetworkX directed graph containing chunks, entities and relationships.</p><p><strong>9. Added entity extraction</strong> — entity_service.py started with simple rules for payment terms and departments and later grew to include invoice concepts.</p><p><strong>10. Added relationship extraction</strong> — relationship_service.py represented business facts such as Immediate → requires_approval → Finance and Net 45 → may_apply_to → Selected Supplier.</p><p><strong>11. Added graph traversal</strong> — graph_retrieval_service.py supported forward and reverse traversal so a question could find a relationship from either side of the edge.</p><p><strong>12. Built hybrid retrieval</strong> — hybrid_retrieval_service.py combined vector results with graph evidence. Exact/normalized entity matching was given priority, with semantic matching as a fallback.</p><p><strong>13. Grounded the answer</strong> — answer_service.py sent the retrieved document context and graph context to GPT-5-mini and returned the answer together with source information.</p><p><strong>14. Added multi-document support</strong> — I indexed Supplier_Policy_Test.pdf and Second_Policy.pdf into the same stores, with document-aware chunk IDs and metadata.</p><p><strong>15. Added duplicate prevention</strong> — document_store.py used SHA-256 file hashes so the exact same document would not be indexed twice.</p><p><strong>16. Added semantic entity matching</strong> — semantic_entity_service.py allowed different wording to find graph concepts that were not an exact string match.</p><p><strong>17. Cached entity embeddings</strong> — entity_embedding_store.py persisted entity embeddings so they were not regenerated for every question.</p><p><strong>18. Expanded the graph</strong> — I added invoice requirements and an invoice-hold relationship so the graph could represent more than payment terms.</p><p><strong>19. Built the React UI</strong> — App.jsx, App.css, main.jsx and api.js turned the backend into a usable assistant showing answers, provenance, graph evidence and retrieval mode.</p><p><strong>20. Evaluated the system</strong> — I tested selected suppliers, strategic suppliers, immediate payment approval, standard payment terms, non-compliant invoices and invoice requirements.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*4zWwheXW1IlqPTmti3b3vQ.png" /><figcaption><em>Figure 4. What happens when a PDF is indexed and what happens when a question is asked.</em></figcaption></figure><h3>The backend, file by file</h3><p>This was the part I most needed as a beginner. I wanted to know not only what the code did, but why each file existed. This became my mental map of the backend.</p><ol><li><strong>main.py — </strong>FastAPI entry point. Defines /health, /upload and /ask, configures CORS, and wires the services together.</li></ol><p>2. <strong>pdf_service.py — </strong>Extracts text from PDF pages and preserves page numbers.</p><p>3. <strong>chunk_service.py — </strong>Splits page text into overlapping chunks. The first version was intentionally simple.</p><p>4. <strong>embedding_service.py — </strong>Creates OpenAI embeddings using text-embedding-3-small.</p><p>5. <strong>vector_store.py — </strong>Stores embeddings in FAISS and stores chunk metadata in JSON.</p><p>6. <strong>document_store.py — </strong>Tracks indexed documents using SHA-256 hashes for duplicate prevention.</p><p>7. <strong>graph_service.py </strong>— Persists the NetworkX knowledge graph and exposes node/edge creation helpers.</p><p>8. <strong>entity_service.py</strong> — Extracts business entities using lightweight rules.</p><p>9. <strong>relationship_service.py — </strong>Turns policy sentences into structured business relationships.</p><p>10. <strong>graph_retrieval_service.py — </strong>Finds graph entities and traverses forward or reverse relationships.</p><p>11. <strong>semantic_entity_service.py — </strong>Embeds richer graph-entity descriptions and finds semantically similar concepts.</p><p>12. <strong>entity_embedding_store.py — </strong>Persists graph-entity embeddings so they can be reused.</p><p>13. <strong>hybrid_retrieval_service.py — </strong>Coordinates vector search, exact entity matching, semantic fallback and graph retrieval.</p><p>14. <strong>indexing_service.py — </strong>The orchestration layer for PDF indexing: extract → chunk → embed → FAISS → entities → relationships → persist.</p><p>15. <strong>answer_service.py — </strong>Builds grounded document/graph context, calls GPT-5-mini and returns answer + provenance + graph evidence.</p><h3>The frontend: where the RAG becomes visible</h3><p>The frontend is intentionally simple. I wanted the UI to show not only the final answer, but how the answer was retrieved.</p><p>· App.jsx — main React component: file upload, question box and answer display.</p><p>· App.css — styles the cards, source evidence, retrieval summary and graph relationships.</p><p>· index.css — global frontend styling.</p><p>· main.jsx — React entry point.</p><p>· frontend/src/services/api.js — calls the FastAPI /upload and /ask endpoints.</p><p>· package.json — React/Vite scripts and frontend dependencies.</p><h3>What happens when a PDF is uploaded?</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*XidH_0RXsDEOQjCA2Qk30A.png" /></figure><p><strong>What happens when a question is asked?</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*-5-zIprfzzshwGA9Ndozog.png" /></figure><blockquote><strong>An important lesson</strong><br> Hybrid RAG does not mean every question must use both retrieval sources. For “What is the standard supplier payment term?”, vector retrieval was sufficient and graph retrieval returned zero useful relationships. For invoice requirements, the graph added six structured relationships.</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/609/1*y-2mOwEUZXzE1JFqLASVtA.png" /></figure><h3>What I learned from the mistakes</h3><p><strong>Chunking can fail quietly. </strong>My first chunking loop could keep returning to the last chunk because the overlap moved the start position backward. The fix was to break when the end of the text was reached.</p><p><strong>Semantic similarity is only a signal. </strong>A question about non-compliant invoices once pulled a Net 60 graph relationship because the semantic match was too broad. The graph itself was correct; retrieval selection was the problem.</p><p><strong>Exact matching can be safer when the entity is explicit. </strong>For “Who must approve immediate payment?”, recognizing the explicit term “Immediate” before semantic fallback produced the correct Finance and Procurement graph evidence.</p><p><strong>More context improves entity embeddings. </strong>Embedding only “Net 45” was weaker than embedding a richer description such as “Net 45 is a payment term that may apply to selected suppliers.”</p><p><strong>Provenance deserves first-class treatment. </strong>Document name, page number and chunk ID became part of the result so users could understand where an answer came from.</p><p><strong>Not every question needs graph evidence. </strong>A robust hybrid design should allow the graph to contribute zero results when the graph has nothing useful to add.</p><h3>Final takeaway</h3><p>What started as a question about “how RAG works” became a practical system that I could inspect, test and improve layer by layer. I started with Vector RAG because it was easier to understand. Then I added persistence, multiple documents, provenance, a knowledge graph, semantic entity matching and hybrid retrieval. Each step answered a different question about how enterprise GenAI systems work.</p><p>The biggest takeaway for me is that understanding GenAI architecture is much easier when the architecture is something you can run. Once I could see a question turn into embeddings, retrieved chunks, graph relationships and a grounded answer, RAG stopped feeling like a collection of buzzwords and started feeling like an engineering workflow.</p><blockquote><strong>If you are also learning RAG</strong><br> Do not start by trying to build the most advanced agentic system. Build one small retrieval pipeline. Then add one capability at a time. Make every layer visible. Test it with real questions. The debugging is part of the learning.</blockquote><h3>Appendix: Backend Build Order</h3><p>For my own reference, I followed the sequence below while building the backend. I found this approach helpful because each step introduced one new concept and built naturally on the previous one.</p><p>StepFilePurpose / Concept</p><p>01backend/requirements.txt Dependencies and environment setup</p><p>02backend/services/pdf_service.pyPDF text extraction</p><p>03backend/services/chunk_service.pyText chunking with overlap</p><p>04backend/services/embedding_service.pyCreating OpenAI embeddings</p><p>05backend/database/vector_store.pyFAISS vector storage and similarity search</p><p>06backend/services/graph_service.pyPersistent NetworkX knowledge graph</p><p>07backend/services/entity_service.pyEntity extraction</p><p>08backend/services/relationship_service.pyRelationship extraction</p><p>09backend/services/graph_retrieval_service.pyGraph entity lookup and traversal</p><p>10backend/services/indexing_service.pyOrchestrating the document indexing process</p><p>11backend/services/answer_service.pyGrounded answer generation and source provenance</p><p>12backend/database/document_store.pySHA-256 based duplicate-document prevention</p><p>13backend/services/hybrid_retrieval_service.pyCombining vector and graph retrieval</p><p>14backend/services/semantic_entity_service.pySemantic matching of graph entities</p><p>15backend/services/entity_embedding_store.pyCaching entity embeddings</p><p>16backend/main.pyFastAPI API layer and application wiring</p><p>#RAG #GraphRAG #HybridGraphRAG #GenerativeAI #GenAI #LLM #VectorSearch #KnowledgeGraph #Embeddings #FAISS #OpenAI #Python #FastAPI #EnterpriseAI #OracleFusion</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/905/1*iU2YY4D2qrtHqdd6_aF1Ow.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/885/1*NMngYeiM7SP4JbDCk80Ozg.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/905/1*6Kt-dF7cEETMjNFKKV403A.png" /></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=da2c740bbbcb" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Oracle Fusion — Data Conversion: HCM / HDL]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/@jhasamiree1/oracle-fusion-data-conversion-hcm-hdl-df8c9a05a53f?source=rss-8a00b14ded6e------2"><img src="https://cdn-images-1.medium.com/max/1774/1*0qZv_yaD7BMLaXv2tKk8SQ.png" width="1774"></a></p><p class="medium-feed-snippet">Hi All,</p><p class="medium-feed-link"><a href="https://medium.com/@jhasamiree1/oracle-fusion-data-conversion-hcm-hdl-df8c9a05a53f?source=rss-8a00b14ded6e------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://medium.com/@jhasamiree1/oracle-fusion-data-conversion-hcm-hdl-df8c9a05a53f?source=rss-8a00b14ded6e------2</link>
            <guid isPermaLink="false">https://medium.com/p/df8c9a05a53f</guid>
            <category><![CDATA[oracle-fusion-hcm]]></category>
            <category><![CDATA[oracle-hcm]]></category>
            <category><![CDATA[hdl]]></category>
            <category><![CDATA[oracle-fusion-technical]]></category>
            <dc:creator><![CDATA[Samir Jha|Oracle Architect (OCI|OIC|SaaS Appl)]]></dc:creator>
            <pubDate>Fri, 02 Oct 2026 14:16:05 GMT</pubDate>
            <atom:updated>2026-10-02T14:16:05.311Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[Oracle Fusion — Data Conversion: Supplier Import]]></title>
            <link>https://medium.com/@jhasamiree1/oracle-fusion-data-conversion-supplier-import-a8bc70bd6a2a?source=rss-8a00b14ded6e------2</link>
            <guid isPermaLink="false">https://medium.com/p/a8bc70bd6a2a</guid>
            <category><![CDATA[data-conversion]]></category>
            <category><![CDATA[supplier-import]]></category>
            <category><![CDATA[oracle-fusion-technical]]></category>
            <dc:creator><![CDATA[Samir Jha|Oracle Architect (OCI|OIC|SaaS Appl)]]></dc:creator>
            <pubDate>Thu, 24 Sep 2026 10:48:35 GMT</pubDate>
            <atom:updated>2026-10-06T07:38:09.815Z</atom:updated>
            <content:encoded><![CDATA[<p>Note: The blog has been co-authored by <a href="https://medium.com/u/1b1d94bd006">Kshemesh Mehta</a></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*LxVOSubT0QwDqU1eBcHDng.png" /><figcaption>Supplier Import</figcaption></figure><p>In this blog, I’ll explain the process of creating supplier in Oracle Fusion.</p><p>Pre-requisites -</p><ol><li>Should be an Employee — Manage Users</li><li>Role — Supplier Manager</li><li>Procurement Agent — Manage Procurement Agents</li><li>Approval Process for Supplier — needs to be setup.</li><li>Data Access for the respective BUs. — Manage Data Access Sets%</li></ol><p>This article provides a practical overview of supplier data conversion in Oracle Fusion, covering the key stages involved in preparing, cleansing, mapping, and importing supplier information. It explains how supplier master data, addresses, sites, contacts, and banking details are organized and loaded using the appropriate import templates. The focus is on understanding the conversion process, avoiding common data errors, and ensuring that supplier records are successfully created and validated in Oracle Fusion.</p><h3>Create Suppliers Using the Import Suppliers Process</h3><p>Oracle Fusion provides several import processes to create new supplier records from external systems and update or delete existing supplier information. These processes support the migration and maintenance of supplier master data, including addresses, sites, contacts, classifications, and related details.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*74PHySUQq-JdpIdhrISUzw.png" /></figure><p>The key supplier import processes include:</p><ul><li>Import Suppliers</li><li>Import Supplier Addresses</li><li>Import Supplier Sites — includes third-party payment relationships</li><li>Import Supplier Site Assignments</li><li>Import Supplier Contacts — includes user accounts and contact addresses</li><li>Import Supplier Business Classifications</li><li>Import Supplier Products and Services Categories</li><li>Import Supplier Attachments</li><li>Purge Supplier Interface Records</li></ul><h3>Supplier Import:OIC Process Explained:</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*K_gtOJ1SZjPWBenyQRnPKQ.png" /></figure><h3>End-to-End Process</h3><p>1. Load Data into Staging Tables (ATP) Extract supplier data from source files and load it into staging tables in Oracle Autonomous Transaction Processing (ATP).</p><p>2. Perform Data Transformation Validate, cleanse, and transform the source data according to Oracle Fusion supplier import requirements.</p><p>3. Upload Files to Oracle Fusion Cloud Generate the required import files, package them into a ZIP file, and upload them to Oracle Fusion Cloud using OIC.</p><p>4. Trigger Supplier Import and Receive Callback Submit the ImportSuppliers ESS job and receive a callback notification upon completion.</p><p>5. Generate Reconciliation Reports Trigger reconciliation reports, receive the output through callback integrations, and review successful and rejected records for further action.</p><h4>Load the data into STG tables:</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*WndlQV9q5JFPcFx7hSGzYA.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/629/1*jkmZJ_aeFdNAKs2UATsvaQ.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*1Vka3yINTlFY3vRA2i_z0A.png" /></figure><h4>Upload to Cloud:</h4><p>In this example, the integration prepares parameters for processing a supplier import file.</p><p>The following variables are configured:</p><p>Variable Purpose:</p><p>p_tnf_path Stores the file transfer or source path.</p><p>v_file_name Defines the input filename, such as PozSuppliersInt.csv.</p><p>filedir Specifies the directory where supplier files are maintained.</p><p>v_zipfile_name Dynamically generates the ZIP filename using the batch identifier.</p><p>v_Job_Package_Name Stores the Oracle ESS job package path.</p><p>v_Job_Def_Name Defines the ESS job to be executed, such as ImportSuppliers.</p><p>v_zip_file_prefix Defines the naming prefix for the generated ZIP file.</p><h3>Generating a Property File Using XSLT in Oracle Integration</h3><p>In enterprise integrations, configuration parameters are often stored in property files to enable dynamic processing and simplify integration maintenance.</p><p>In this example, an XSLT transformation is used to generate a structured property file containing ESS job details and runtime parameters for a supplier import process.</p><h3>XSLT Mapping</h3><pre>&lt;xsl:template match=&quot;/&quot; xml:id=&quot;id_11&quot;&gt;<br>    &lt;nstrgmpr:Write xml:id=&quot;id_12&quot;&gt;<br>        &lt;ns21:SupplierProprecset xml:id=&quot;id_28&quot;&gt;<br>            &lt;ns21:SupplierProprec xml:id=&quot;id_29&quot;&gt;</pre><pre>                &lt;ns21:JOB_PKG_NAME xml:id=&quot;id_47&quot;&gt;<br>                    &lt;xsl:value-of select=&quot;$v_Job_Package_Name&quot;/&gt;<br>                &lt;/ns21:JOB_PKG_NAME&gt;</pre><pre>                &lt;ns21:JOB_DEF_NAME xml:id=&quot;id_49&quot;&gt;<br>                    &lt;xsl:value-of select=&quot;$v_Job_Def_Name&quot;/&gt;<br>                &lt;/ns21:JOB_DEF_NAME&gt;</pre><pre>                &lt;ns21:ZIP_FILE_PREFIX xml:id=&quot;id_51&quot;&gt;<br>                    &lt;xsl:value-of select=&quot;$v_zip_file_prefix&quot;/&gt;<br>                &lt;/ns21:ZIP_FILE_PREFIX&gt;</pre><pre>                &lt;ns21:PARAM1 xml:id=&quot;id_30&quot;&gt;<br>                    &lt;xsl:value-of select=&quot;&#39;NEW&#39;&quot;/&gt;<br>                &lt;/ns21:PARAM1&gt;</pre><pre>                &lt;ns21:PARAM2 xml:id=&quot;id_70&quot;&gt;<br>                    &lt;xsl:value-of select=&quot;&#39;N&#39;&quot;/&gt;<br>                &lt;/ns21:PARAM2&gt;</pre><pre>                &lt;ns21:PARAM3 xml:id=&quot;id_72&quot;&gt;<br>                    &lt;xsl:value-of<br>                        select=&quot;/nssrcmpr:execute/ns15:request-wrapper/ns15:BatchId&quot;/&gt;<br>                &lt;/ns21:PARAM3&gt;</pre><pre>            &lt;/ns21:SupplierProprec&gt;<br>        &lt;/ns21:SupplierProprecset&gt;<br>    &lt;/nstrgmpr:Write&gt;<br>&lt;/xsl:template&gt;</pre><figure><img alt="" src="https://cdn-images-1.medium.com/max/529/1*ahPrCjWsC_JVkB-5yKk64w.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/530/1*WOrzDpJEv0Sp47U6xxp_hg.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/523/1*kip47Cqf7_wYWEzbAVl6Lg.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/528/1*kbL9vX-1SKWMR16a3UL3nw.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*lb1Bn_HpSz0wbRX-Xq8dJQ.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*KCVGmy9Og0Y8bky2YBXtPA.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*IA4VCulGN9Fpwn-EKlfJxg.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*I3R4pPM5ooO_dCvwRhUOEg.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/1*dbtBYfX1AJnTRksNj8iGPA.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*hf4er9inJiDGmKnyy1a7Dw.png" /></figure><h4>Callback OIC:</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*2rJr3jsrXjwKFvsuIZgk6Q.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/540/1*Mrge7_G2JzY_iwtQZo0UTA.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*O4pGk6LcFwkXkOyo8OQWiw.png" /></figure><p><strong>Reconciliation2 Call:</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*SR9QejvLapnV9uK5JwTm8A.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*2CJyvRlgBEBG0CoHKj7dPw.png" /></figure><p><strong>Recon2 Call back:</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*MDk5WxJlFP6AOgBC0tjFcw.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*gr7_dnCCtf0CUdY3ok-tjg.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*AsShDO_LL1LvxrIPOdgimg.png" /></figure><h4>Transformation Report :</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*bpxO4KC2IruLfu4au7-tkQ.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/634/1*zLEgCgR17NpeSaTCnSJS-g.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*znLBIOQuE9vWIckRXYPfJg.png" /></figure><p>The below .zip would appear in UCM..</p><p>Files starting with ImportBulkData are used in callback mechanism</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/374/1*6LZzzrSLvZsQvR3-HYstYw.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/728/1*9D32EF4dQoQwHIpIozJT3A.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*kD39onJj7lw491HW8v7U1A.png" /></figure><p><a href="https://docs.oracle.com/en/cloud/saas/procurement/25c/oaprc/create-suppliers-with-the-import-suppliers-process.html#s20030290">Using Procurement</a></p><p>Properties File:</p><p>/oracle/apps/ess/prc/poz/supplierImport/,ImportSupplierAddresses,PozSupplierAddressesInt,NEW,N,1771</p><p>/oracle/apps/ess/prc/poz/supplierImport/,ImportSupplierSites,PozSupplierSitesInt,NEW,N,1728</p><p>/oracle/apps/ess/prc/poz/supplierImport/,ImportSupplierContacts,PozSupContactsInt,NEW,N,1787</p><p>After Supplier Import, in Redwood theme..it looks like as below:</p><p>There are 9 tabs:</p><p>Organization Details , Addresses , Contacts, Sites, Transaction Tax, Payments, Biz Classifications, Products and Services, Qualifications</p><p>Profile Option Setup — to restrict emails..</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*JwsLPhPTANV_gyeNHppCjQ.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*rYbFvaRVTcxc1sFiyJSDNQ.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*N6ZSZKox9lvJKnWOacF8lw.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*0GJ5VuoRvYbB_B87OpxlBg.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*DjUOEiZ2B6Jce50Mek3ZXQ.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*DBHC1n6D7GdNAQxfu3PMdQ.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*dc0jlJQM1e31JFVUawTC6Q.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*yXY2TwW7VItzdRH61Cqk3g.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*bStHgIzJDo6wjmLKhHJU1Q.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*YVzJYLl8eZhr3JpKmHJEEw.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*xVDAx_-uFFe9fs7ugrFvcw.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*86XSk3DB6iCtgUTrexgNmg.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*eQ-NCsp2wzuepLePJbdVrQ.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ehOVbZvBjdV0FlboidXY-g.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ImPsuj7hmx086y8VbNpKWg.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*OIN3lVcqx5sndE_2l8nNNg.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*spONUwKrVxwoqmFPcndavw.png" /></figure><p>#OracleFusion #OracleIntegrationCloud #OIC #SupplierImport #SupplierConversion #OracleSCM #DataConversion #OracleERP #CloudIntegration #FBDI #OracleESS #EnterpriseIntegration #OracleTechnical #ERPImplementation</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a8bc70bd6a2a" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[A practical journey with React, Python, FastAPI, FAISS and OpenAI — from uploading a PDF to…]]></title>
            <link>https://medium.com/@jhasamiree1/a-practical-journey-with-react-python-fastapi-faiss-and-openai-from-uploading-a-pdf-to-c1d2d68d4e0d?source=rss-8a00b14ded6e------2</link>
            <guid isPermaLink="false">https://medium.com/p/c1d2d68d4e0d</guid>
            <dc:creator><![CDATA[Samir Jha|Oracle Architect (OCI|OIC|SaaS Appl)]]></dc:creator>
            <pubDate>Sun, 06 Sep 2026 14:09:54 GMT</pubDate>
            <atom:updated>2026-09-06T14:10:22.624Z</atom:updated>
            <content:encoded><![CDATA[<h3>A practical journey with React, Python, FastAPI, FAISS and OpenAI — from uploading a PDF to retrieving the right page and generating a grounded answer.</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*UeVL2NtGZTpG9NzOSGJAlA.png" /></figure><h3>1. Why I decided to build a small RAG application</h3><p>I had been reading about <strong>RAG, embeddings, vector databases, and grounding</strong>, but I found that these concepts started making much more sense once I actually tried building something myself.</p><p>Instead of creating another generic chatbot, I decided to build an application — with the help of AI and Codex — that was closer to a real-world enterprise use case: <strong>asking questions about a supplier policy document</strong>.</p><p>The idea was simple. A user uploads a <strong>Supplier Policy PDF</strong>, the application <strong>indexes the document</strong>, and the user can then ask questions such as:</p><blockquote>“What are the standard supplier payment terms?”</blockquote><p>Behind the scenes, the application searches the document for the most relevant information, retrieves the appropriate content, and sends that context along with the user’s question to the <strong>LLM</strong>.</p><p>The application then generates an answer and, importantly, provides the <strong>source document and page number</strong> where the information was found.</p><p>Building this small application helped me move from simply reading about RAG concepts to actually understanding how the different pieces work together in a real application.</p><h3>2. What I wanted the final application to do</h3><p>I deliberately kept the first version of the application simple and focused on the core RAG workflow. The application was designed to allow a user to:</p><ul><li>Upload a <strong>Supplier Policy PDF</strong>.</li><li>Extract the text from the PDF.</li><li>Split the extracted text into smaller, manageable chunks.</li><li>Create an embedding for each chunk.</li><li>Store the embeddings in <strong>FAISS</strong>, along with relevant metadata such as the document name and page number.</li><li>Ask a question about the document.</li><li>Create an embedding for the user’s question.</li><li>Search <strong>FAISS</strong> to find the most relevant chunks.</li><li>Use the retrieved content to <strong>ground the LLM’s response</strong>.</li><li>Display the final answer along with the <strong>source document and page number</strong>.</li></ul><p>The goal was to keep the application small enough to understand each step clearly, while still building a complete end-to-end RAG pipeline.</p><h3>3. Technology stack</h3><p>For this project, I used <strong>React and JavaScript</strong> to build the frontend, while <strong>Python</strong> handled the core RAG logic. <strong>FastAPI</strong> was used to create the backend APIs and connect the frontend with the RAG pipeline.</p><p>For the AI components, I used <strong>OpenAI’s </strong><strong>text-embedding-3-small</strong> model to convert document text into embeddings, <strong>FAISS</strong> to perform vector similarity searches, and <strong>GPT-5-mini</strong> to generate answers based on the retrieved document context.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/958/1*qAAbarSEfY5yTPWObWqokg.png" /></figure><h3>4. The architecture</h3><p>Before writing the code, I simplified the application into <strong>two main phases: Indexing and Question Answering</strong>.</p><h3>Phase 1: Indexing the document</h3><p>The first phase happens when the user uploads the Supplier Policy PDF. The application extracts the text from the document, splits it into smaller chunks, and converts each chunk into an embedding. These embeddings, along with metadata such as the document name and page number, are stored in <strong>FAISS</strong>. Once this process is complete, the document is ready to be searched.</p><h3>Phase 2: Answering a question</h3><p>When a user asks a question, the application converts the question into an embedding and uses it to perform a similarity search in FAISS. The most relevant document chunks are retrieved and provided to the LLM as context. This process, known as <strong>grounding</strong>, helps the LLM generate an answer based on the actual content of the document. The application then displays the answer along with the relevant source and page number.</p><p>The diagram above shows how these two phases work together to form the complete RAG pipeline.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/931/1*OoOv4NEhMkQL7zyukRQEzA.png" /></figure><h3>Step 1 — Create the project</h3><p>I created separate frontend and backend folders. This made the application easier to understand because I could clearly see which part belonged to the browser and which part belonged to the Python backend.</p><p>supplier-policy-rag/<br> ├── backend/<br> └── frontend/</p><h3>Step 2 — Set up Python</h3><p>Inside the backend folder, I created a virtual environment. It keeps this project’s Python packages isolated from the rest of my machine.</p><p>python -m venv .venv<br> .\.venv\Scripts\Activate.ps1</p><p>After activating it, I installed FastAPI, Uvicorn, python-multipart, python-dotenv, OpenAI, FAISS, PyPDF and NumPy.</p><h3>Step 3 — Configure the OpenAI key</h3><p>I kept the OpenAI API key on the backend in a .env file rather than putting it into React. This is important because anything shipped to the browser should be treated as visible to the user.</p><p>backend/.env<br> <br> OPENAI_API_KEY=your_api_key_here</p><p>React calls my backend. The backend is the part that talks to OpenAI.</p><h3>Step 4 — Create the FastAPI backend</h3><p>The backend exposes three simple endpoints:</p><p>GET /health<br> POST /index<br> POST /ask</p><p>I think of FastAPI as the bridge between the React screen and the Python RAG engine. React handles the user interface, while FastAPI receives the PDF and questions and passes them into the RAG workflow.</p><h3>Step 5 — Extract text from the PDF</h3><p>When the user selects Supplier_Policy_Test.pdf and clicks Index PDF, the backend reads the document page by page. I intentionally keep the page number with the extracted text.</p><p>{<br> “page”: 4,<br> “text”: “Standard payment term: Net 30 days…”<br> }</p><p>Keeping the page number is what later allows the application to say where the answer came from instead of returning a source-free answer.</p><h3>Step 6 — Chunk the document</h3><p>I don’t want to treat the entire PDF as one giant piece of text. I split it into smaller chunks so retrieval can focus on the part of the policy that is actually relevant.</p><p>Chunk ID: 101<br> Document: Supplier_Policy_Test.pdf<br> Page: 4<br> Text: Standard payment term: Net 30 days…</p><p>The exact chunking strategy can be tuned later. For this first application, I kept it simple and focused on understanding the pipeline.</p><h3>Step 7 — Create embeddings</h3><p>This was one of the concepts I wanted to understand rather than simply use as a black box. An embedding model converts text into a numerical vector. Texts with related meanings tend to have vectors that are closer together in vector space.</p><p>I used text-embedding-3-small. With its default configuration, the vector in this application has 1536 dimensions.</p><p>Text<br> “Standard payment term is Net 30 days…”<br> ↓<br> text-embedding-3-small<br> ↓<br> [0.012, -0.034, 0.087, …]<br> 1536 values</p><p>The 1536 number is not a universal rule for embeddings. It is the default size for this model configuration. Other models or configurations can use different dimensions.</p><h3>Step 8 — Store the vectors in FAISS</h3><p>Once the chunks have become vectors, I need a way to search them. That is where FAISS comes in. FAISS stands for Facebook AI Similarity Search and is designed for efficient similarity search over vectors.</p><p>One distinction became very clear to me: FAISS is not the original PDF. The vector index represents the numerical vectors, while metadata maps a vector back to its document, page and text.</p><p>FAISS<br> Vector ID 0 → 1536-dimensional vector<br> Vector ID 1 → 1536-dimensional vector<br> Vector ID 2 → 1536-dimensional vector<br> <br> Metadata<br> Vector ID 0 → PDF + page + text<br> Vector ID 1 → PDF + page + text<br> Vector ID 2 → PDF + page + text</p><h3>Step 9 — Indexing is complete</h3><p>After extraction, chunking, embedding and vector indexing, the document is ready for questions. In my test application, the UI displayed: ‘Indexed Supplier_Policy_Test.pdf: 8 pages, 8 chunks.’ That was a useful confirmation that the indexing phase had completed.</p><h3>Step 10 — Ask a question</h3><p>Now the user can ask: ‘What are the supplier payment terms?’ The question goes through the same embedding process as the document chunks.</p><p>Question<br> “What are the supplier payment terms?”<br> ↓<br> Embedding model<br> ↓<br> Question vector<br> ↓<br> FAISS</p><h3>Step 11 — Similarity search</h3><p>FAISS compares the question vector with the vectors created from the document. The aim is to find the chunks whose meaning is most relevant to the question. A payment question should bring back the payment section rather than an unrelated supplier-registration section.</p><h3>Step 12 — Ground the LLM</h3><p>This is probably the most important RAG concept in the whole application. Retrieval finds relevant information, but retrieval itself does not write the final answer.</p><p>I take the user’s question and the retrieved policy content and provide them to the LLM as context. Making the response rely on that retrieved source is commonly called grounding.</p><p>Question<br> +<br> Retrieved policy context<br> ↓<br> LLM<br> ↓<br> Grounded answer</p><p>The practical idea is: instead of asking the model to answer only from its general knowledge, I give it the relevant company information at the time of the question.</p><h3>Step 13 — Return the answer with the source</h3><p>For the payment-terms question, the application returned an answer based on the policy and showed Supplier_Policy_Test.pdf, Page 4 as the source. The UI also displayed the retrieved source snippet and similarity score.</p><p>That source reference matters in an enterprise setting. Users are much more comfortable with an AI answer when they can see which document and page were used.</p><h3>Step 14 — Test the embedding separately</h3><p>Before relying on the full application, I tested the embedding connection separately. With the environment loaded, I created an embedding for the word ‘hello’ and printed its length.</p><p>1536</p><p>That small test helped me separate API/configuration problems from problems in the rest of the application.</p><h3>Step 15 — Start the application</h3><p>To start the backend, I open PowerShell, activate the environment and run:</p><p>cd …\supplier-policy-rag\backend<br> .\.venv\Scripts\Activate.ps1<br> uvicorn main:app — reload — port 8000</p><p>Uvicorn is the server that runs the FastAPI application. FastAPI defines the API; Uvicorn serves it. I can check the API using the Swagger page at <a href="http://127.0.0.1:8000/docs.">http://127.0.0.1:8000/docs.</a></p><p>In a second PowerShell window, I start the React frontend:</p><p>cd …\supplier-policy-rag\frontend<br> npm run dev</p><p>The React development server then provides the browser UI, typically at <a href="http://localhost:5173.">http://localhost:5173.</a></p><h3>Step 15.1 — A small issue I encountered — CORS</h3><p>Because the React frontend and FastAPI backend run on different ports, the browser treats them as different origins. During development I hit a CORS error when React tried to call the /index endpoint.</p><p>The useful lesson was that a CORS error does not necessarily mean the RAG logic is broken. The backend needs to explicitly allow the frontend origin during development.</p><h3>Step 16— Testing the application</h3><p>Once both servers are running, I open the React application, choose Supplier_Policy_Test.pdf, and click Index PDF. After indexing, I ask questions such as:</p><p>· What are the standard supplier payment terms?</p><p>· When is immediate payment allowed?</p><p>· Who needs to approve immediate payment?</p><p>· Can strategic suppliers receive Net 60 terms?</p><p>· What information must be included on a supplier invoice?</p><p>· What happens when an invoice does not meet the policy requirements?</p><p>I also like to test a question that is clearly outside the document. If the information is not present, the application should avoid confidently inventing an answer. Ideally it should say that the information could not be found in the supplier policy.</p><h3>Step 17— What each component is really doing</h3><p>· React is the screen the user interacts with.</p><p>· FastAPI exposes the Python backend as APIs.</p><p>· Uvicorn runs the FastAPI application.</p><p>· PDF processing extracts the source content.</p><p>· Chunking breaks documents into searchable pieces.</p><p>· Embeddings turn text into vectors that capture semantic relationships.</p><p>· FAISS searches those vectors.</p><p>· Metadata tells us which document and page a vector came from.</p><p>· Grounding gives retrieved information to the LLM as context.</p><p>· The LLM turns that context into a natural-language answer.</p><h3>Step 19 — Where I would take the application next</h3><p>The first version is intentionally small. The next step would be to make it more enterprise-oriented. Instead of one supplier policy PDF, I could index multiple policies, contracts, invoice documents and Oracle documentation.</p><p>I could also connect live Oracle Fusion data through an API or tool. For example, the system could retrieve the supplier policy through RAG and retrieve the actual payment terms from a purchase order. The LLM could then compare the two and explain whether the PO follows the policy.</p><p>Supplier Policy<br> │<br> └── RAG → policy rules<br> <br> Oracle Fusion<br> │<br> └── API/tool → live PO data<br> <br> ↓<br> <br> LLM<br> ↓<br> Policy comparison</p><h3>Step 20— RAG, agents and MCP</h3><p>Building this project also helped me understand how <strong>RAG, AI Agents and MCP can work together</strong> in a real-world enterprise application.</p><p>In the current application, <strong>RAG is responsible for retrieving information</strong> from the Supplier Policy PDF. When a user asks a question, the system searches the indexed document, retrieves the most relevant content and provides it to the LLM. The LLM then generates an answer based on that retrieved information.</p><p>However, many real business questions require information from more than just a document.</p><p>For example, imagine a user asks:</p><blockquote><strong><em>“Check PO 40683 and tell me whether its payment terms comply with the Supplier Policy.”</em></strong></blockquote><p>Answering this question requires information from multiple sources:</p><ul><li>The <strong>Supplier Policy</strong> needs to be searched to find the approved payment terms.</li><li><strong>Oracle Fusion</strong> needs to be queried to retrieve the actual payment terms on PO 40683.</li><li>In some scenarios, a <strong>SQL database</strong> may need to be queried for additional transaction or supplier information.</li></ul><p>This is where an <strong>AI Agent</strong> becomes useful.</p><p>Unlike a basic RAG application, an AI Agent can analyze the user’s request and decide <strong>which source or tool it needs to use</strong>. It can retrieve information from RAG, call an API, execute a database query and then combine the results before asking the LLM to generate the final response.</p><p><strong>MCP (Model Context Protocol)</strong> can help standardize how an AI application connects to these external tools and data sources. Instead of building completely different integration mechanisms for every tool, MCP provides a structured approach for exposing capabilities such as APIs, databases and enterprise applications to AI systems.</p><p>A future version of this application could therefore evolve from a simple document-based RAG application into an intelligent enterprise assistant.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*E0kbD19-KRGWrI_gzrndCQ.png" /></figure><h3>Step 22— The biggest thing I learned</h3><p>The <strong>biggest lesson</strong> for me was that <strong><em>RAG </em></strong>is not just about adding a vector database to an LLM application. There are several separate steps, and each one solves a different problem.</p><p>The document has to be processed correctly. The chunks have to retain useful context. The embeddings need to represent the content well. Retrieval needs to find the right chunks. Finally, the LLM needs to be grounded with the retrieved information.</p><p>Once I saw the complete flow working, the individual terms stopped feeling like separate AI buzzwords. They became parts of one pipeline.</p><h3>Final takeaway</h3><p>The application started with a simple question: ‘Can I ask questions about a supplier policy PDF?’ The answer turned into a complete RAG pipeline:</p><p>For me, building this small application was much more useful than only reading about RAG. It showed exactly where embeddings fit, why a vector store is needed, what retrieval actually means, and why grounding matters.</p><p>More importantly, it showed how the same basic idea can grow into something bigger: an enterprise assistant that combines company documents, Oracle Fusion data, SQL, APIs and tools to answer real business questions.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*7iv8TlUWh69mz-kvw_iDaQ.png" /></figure><h3><strong>Screen shots :</strong></h3><p><strong>Backend Application:</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*UUnP3pxqqliQbpwZBOy1qw.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*n-3JVg2ctOYpoX52PFeB1A.png" /></figure><p><strong>Front end:</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*LtiVZuib0a6vIMTbKEJ1IA.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*FK6A5_ZmXa260KFynAwC-Q.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*5Q2LwwCvEeQ8d3wbadr0Mw.png" /></figure><p>If you would like to explore the project in more detail, I have shared the complete source code on GitHub. This includes the React frontend, FastAPI backend, PDF indexing process, embeddings, FAISS vector search, and the question-answering flow.</p><p><a href="https://github.com/samirkol/RAG">Explore the complete project on GitHub</a></p><p>#RAG #GenerativeAI #ArtificialIntelligence #LLM #OpenAI #VectorDatabase #FAISS #Embeddings #Python #FastAPI #React #AIEngineering #MachineLearning</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c1d2d68d4e0d" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Oracle Fusion AI Studio — Certified Developer Professional Guide — Part V]]></title>
            <link>https://medium.com/@jhasamiree1/preparing-for-oracle-fusion-ai-agent-studio-certified-developer-professional-rel-26-2-a-47aada95f536?source=rss-8a00b14ded6e------2</link>
            <guid isPermaLink="false">https://medium.com/p/47aada95f536</guid>
            <dc:creator><![CDATA[Samir Jha|Oracle Architect (OCI|OIC|SaaS Appl)]]></dc:creator>
            <pubDate>Sun, 06 Sep 2026 13:00:30 GMT</pubDate>
            <atom:updated>2026-09-24T10:05:26.637Z</atom:updated>
            <content:encoded><![CDATA[<p><em>A practical preparation guide covering security, prompts, topics, tools, testing, tracing, Guided Journeys, integrations, and governance</em></p><p>Preparing for a certification is often more effective when we focus on understanding the concepts behind the questions rather than simply memorizing answers.</p><p>While preparing for the Oracle Fusion AI Agent Studio Certified Developer Professional — Rel 26–2 certification, I came across several scenarios covering Agent Teams, security, prompts, Topics, Tools, Guided Journeys, testing, tracing, integrations, and governance.</p><p>What I found interesting was that many of these questions were connected. Once the underlying concepts were grouped together, the answers became much easier to understand and remember.</p><p>This article organizes those concepts into practical sections so that it can serve both as a learning guide and as a preparation resource. Instead of looking at the questions one by one, the focus is on understanding why a particular Oracle AI Agent Studio component is used in a particular situation.</p><p><strong>1. Agent Security Starts with the Agent Team</strong></p><p>One of the first and most important concepts is understanding where security is configured.</p><p>When an AI agent needs Oracle Fusion role-based security, the important component is the Agent Team.</p><p>An Agent Team is not just a logical grouping of agents. It is also an important part of controlling who can access and interact with agents.</p><p>The preparation scenarios repeatedly reinforce the following ideas:<br>• Oracle Fusion role-based security is associated with the Agent Team.<br>• Custom roles can be configured with permission groups.<br>• Relevant Oracle Fusion duties and functional security policies can be assigned according to the business area.<br>• Different administrators can receive access based on their responsibilities.</p><p>For example, an organization may have agents for HR, SCM, and Finance. It would not be appropriate to give every administrator the same level of access. An HR specialist should manage HR-related agents, an SCM administrator should manage SCM-related agents, and end users should receive only the access required to interact with the published agent.</p><p>The underlying principle is simple: access should be based on responsibility, not convenience.</p><p><strong>2. Topics Are How You Organize Business Conversations</strong></p><p>Another major concept is the use of Topics. Topics help define the business areas that an agent is expected to handle.</p><p>Suppose an HR agent is expected to answer questions about Payroll, Leave, and Benefits. A good design is to create separate topics for each area rather than putting everything into one large topic.</p><p>Separate Topics are particularly useful when business areas have different instructions, policies, response styles, or security requirements. The same principle applies to regulatory compliance, tax, audit, multiple languages, regional policies, escalation logic, and different response styles.</p><p>A useful rule to remember is: if two business areas require different instructions or different business guidance, they are strong candidates for separate Topics.</p><p>It is also important to distinguish between Topics and Documents. A Topic organizes what the agent is expected to handle. A Document Tool provides reference information that the agent can retrieve. They are related, but they solve different problems.</p><p><strong>3. A Good System Prompt Should Be Clear, Structured, and Actionable</strong></p><p>One of the strongest themes in the preparation material is prompt design.</p><p>When an agent produces confusing or inconsistent answers, the solution is usually not to keep adding more words to the prompt. Instead, the prompt should become more structured.</p><p>A practical approach is to break the system prompt into concise, step-by-step instructions. For example:</p><p>1. Identify the user’s request.<br>2. Determine the relevant business topic.<br>3. Retrieve information from the appropriate source.<br>4. Provide a concise response.<br>5. If required information is unavailable or ambiguous, ask for clarification.<br>6. Do not provide information outside the agent’s authorized scope.</p><p>Clearly labeled sections can also improve prompt readability and maintainability. For example, a prompt can be organized into ROLE, INSTRUCTIONS, INPUT, OUTPUT, and RESTRICTIONS.</p><p>The key lesson is that a well-structured prompt is easier to understand, maintain, and evaluate.</p><p><strong>4. Prompt Engineering Is Not Just About Adding More Instructions</strong></p><p>A common mistake when designing AI agents is assuming that inconsistent answers can always be fixed by making the prompt longer. That is not always the right solution.</p><p>Relevant examples can help clarify expected behavior. Separate personas or targeted instructions may be useful when the same agent serves different audiences, such as employees and managers. Most importantly, ambiguity should be handled explicitly.</p><p>A practical instruction is: If the user’s request is ambiguous, ask a clarifying question before retrieving or presenting information.</p><p>This is especially important when an agent deals with sensitive or business-critical information. The goal is not simply to generate an answer. The goal is to generate an appropriate answer.</p><p><strong>5. Choosing the Right Tool for the Right Job</strong></p><p>One of the most important things to understand in Oracle AI Agent Studio is that tools are not interchangeable. Each tool exists for a different purpose.</p><p>Business Object Tool<br>Use the Business Object Tool when the agent needs structured Oracle Fusion business data, such as inventory, employee information, transactions, or other business records.</p><p>Document Tool<br>Use the Document Tool when the agent needs to retrieve information from policies, manuals, reference material, or other approved documents. Documents should be processed appropriately before they are used for retrieval.</p><p>External REST Tool<br>Use the External REST Tool when the agent needs to interact with an external API. Key configuration considerations include the endpoint URL, HTTP operation, authentication details, and secure integration practices.</p><p>Deep Link Tool<br>Use a Deep Link Tool when the agent needs to navigate a user directly to a relevant Oracle Fusion application page or work area.</p><p>Context and User Session Tools<br>These tools are useful when the agent needs information such as session context, user-related information, conversation state, or values collected during a multi-step interaction.</p><p><strong>6. Guided Journeys: Bringing Agents into the User Experience</strong></p><p>A technically correct AI agent is not useful if users cannot easily access it. This is where Guided Journeys and homepage integration become important.</p><p>A typical implementation flow is:<br>Create Agent → Create Guided Journey → Associate the Journey with the Agent → Configure Homepage Properties → Add the Guided Journey code → Preview → Activate.</p><p>Activation is required for the Guided Journey to become available, while preview is useful for validating the experience before production use.</p><p>A useful troubleshooting lesson is that an agent can be correctly configured but still not be visible to users if the Guided Journey state, homepage properties, or integration configuration is incorrect.</p><p><strong>7. Testing AI Agents: Correctness, Latency, and User Experience</strong></p><p>Testing is another major area. There is no single metric that defines a good AI agent. A strong evaluation considers multiple dimensions.</p><p>Offline Testing<br>Offline testing is useful for evaluating agent responses against reference or expected answers. It provides a controlled way to measure correctness.</p><p>A/B Testing<br>A/B testing compares different configurations or versions and can help evaluate user experience, correctness, latency, and other outcomes.</p><p>Correctness is not the same as speed. Latency and token usage help evaluate efficiency and cost-performance, while correctness evaluates whether the response is semantically aligned with the expected answer.</p><p>The best configuration is not automatically the fastest or the most accurate one. It is the configuration that best fits the required business outcome.</p><p><strong>8. Agent Tracing and Prompt Tracing</strong></p><p>Tracing is essential when an agent does not behave as expected.</p><p>Agent Tracing focuses on the broader agent execution, helping us understand the steps performed by the agent, topic selection, tool calls, and execution flow.</p><p>Prompt Tracing focuses more specifically on prompt execution details and helps analyze what happened during the processing of an individual prompt.</p><p>This distinction is useful during troubleshooting. If the question is, ‘Which step in the agent workflow caused the problem?’ Agent Tracing is useful. If the question is, ‘Why did this particular prompt produce this output?’ Prompt Tracing is more relevant.</p><p><strong>9. Improving Agent Performance: Start with the Trace</strong></p><p>When an agent becomes slow or inconsistent, the best first step is not necessarily to use a larger model. Start by understanding the actual execution.</p><p>Review the trace for redundant tool calls, repeated retrieval, or unexpected execution paths. Review document retrieval for issues related to document sets and chunking. Review prompts and Topics when responses are unnecessarily long or inconsistent.</p><p>Do not optimize blindly. First identify where the inefficiency is happening. The problem may be related to prompt design, tool selection, retrieval configuration, document chunking, or unnecessary execution steps.</p><p><strong>10. APIs and Integrations Must Be Designed with Security</strong></p><p>Enterprise AI agents often access sensitive information such as payroll, employee records, financial data, and inventory. Security therefore cannot be an afterthought.</p><p>A strong integration design should consider HTTPS, appropriate authentication, encryption, logging, role-based access, and least privilege.</p><p>The agent should receive only the permissions required to perform its intended work. External connections should be secured, access should be authenticated, and sensitive actions should be appropriately controlled.</p><p><strong>11. Prompt Guardrails Are Not a Replacement for Security</strong></p><p>Prompt instructions and security controls are not the same thing.</p><p>Telling an agent ‘Do not show payroll information’ is a behavioral instruction. It should not be treated as the only security mechanism if the agent technically has unrestricted access to payroll data.</p><p>A stronger design uses layered security:<br>• Role-based access controls<br>• API authentication<br>• Data authorization<br>• Prompt guardrails<br>• Logging and monitoring</p><p>Security should be enforced by the system architecture, not only by the agent’s instructions.</p><p><strong>12. Structured Output: Tell the Agent Exactly What Is Expected</strong></p><p>When an agent’s response is consumed by a downstream automation system, the output contract must be explicit.</p><p>If the downstream system expects valid JSON, the prompt should clearly state that the response must be valid JSON and should prohibit additional formatting that could break parsing, such as Markdown, backticks, or explanatory text outside the required structure.</p><p>Humans may appreciate additional explanation, but a downstream parser may fail because of it. When AI output becomes input to another system, precision matters.</p><p><strong>13. Session Context Makes Multi-Step Experiences Possible</strong></p><p>Some business processes require multiple pieces of information collected over time. Benefits enrollment is a good example. A user may provide dependent information, select a health plan, and then continue with coverage preferences.</p><p>The agent should not forget information after every step. Relevant values can be stored as session variables so that they remain available during the selected interaction.</p><p>This is a key difference between a chatbot answering isolated questions and an AI agent supporting a business process. The second experience requires context.</p><p><strong>14. Regional and Language Requirements Should Influence Topic Design</strong></p><p>A global organization may need an agent to support multiple countries, languages, and regional policies.</p><p>When policies or guidance differ by region or language, the Topic design should reflect those differences. A single generic Topic containing all regional policies can become difficult to manage and may increase the risk of providing the wrong guidance.</p><p><strong>15. Compliance and Regulatory Agents Need Centralized Guidance</strong></p><p>Compliance scenarios highlight the importance of consistency.</p><p>When multiple agents need to answer questions using common regulatory guidance, centralizing that guidance can help avoid duplication and inconsistency. At the same time, distinct domains such as regulatory, tax, and audit can remain separate Topics when their instructions and business requirements differ.</p><p>A useful principle is: centralize common guidance and separate distinct business responsibilities.</p><p><strong>16. Use Pre-Built Capabilities Where They Fit</strong></p><p>The preparation material also highlights pre-built functionality, such as the My Team Business Object Tool and its pre-built function for searching workers to query team members.</p><p>A good development approach is to first check whether an appropriate pre-built capability already exists. If it fits the requirement, use it. Create a custom integration only when there is a genuine business need that is not already addressed.</p><p><strong>17. A Complete Example: Designing an HR AI Agent</strong></p><p>Let us bring the concepts together with an HR AI Agent that helps employees with Payroll, Leave, Benefits, and Payslip access.</p><p>Topics:<br>• Payroll<br>• Leave<br>• Benefits</p><p>Tools:<br>• Business Object Tool for structured Oracle Fusion data<br>• Deep Link Tool for navigation to relevant pages<br>• User Session capability for multi-step interactions<br>• Document Tool for approved policy reference information where appropriate</p><p>Security:<br>• Agent Team<br>• Permission groups<br>• Relevant roles and duties<br>• Role-based access controls</p><p>Testing and operations:<br>• Offline testing for correctness against expected answers<br>• A/B testing for comparing configurations and outcomes<br>• Tracing for troubleshooting</p><p>User experience:<br>• Guided Journey<br>• Homepage integration<br>• Preview and activation</p><p><strong>18. The Most Important Patterns to Remember</strong></p><p>Security: Agent Team, permission groups, relevant roles and duties, functional security policies, and least privilege.</p><p>Prompt Design: concise step-by-step instructions, clearly labeled sections, relevant examples, clarification rules, and explicit output requirements.</p><p>Topics: separate Topics when business domains, instructions, response styles, regions, or languages differ.</p><p>Business Object Tool: use for structured Oracle Fusion business data.</p><p>Document Tool: use for retrieving information from processed reference documents.</p><p>External REST Tool: use for external APIs with appropriate endpoint, HTTP operation, authentication, and security.</p><p>Deep Link Tool: use when the agent needs to navigate the user directly to a relevant application page.</p><p>Testing: Offline testing focuses on expected or reference answers. A/B testing compares configurations and outcomes. Latency and token usage evaluate efficiency and cost, while correctness evaluates semantic alignment.</p><p>Tracing: Agent Tracing helps analyze broader execution. Prompt Tracing helps analyze individual prompt execution.</p><p><strong>Final Thoughts</strong></p><p>The most valuable lesson from these Oracle AI Agent Studio preparation scenarios is that successful AI agents are built from multiple connected layers.</p><p>A good agent is not simply an LLM plus a prompt. It combines prompts, Topics, Tools, security, testing, tracing, and a well-designed user experience.</p><p>A useful way to think about an AI agent is to ask:</p><p>1. What business domain is the agent handling?<br>2. How should that domain be organized into Topics?<br>3. What data or action does the agent need?<br>4. Which Tool is appropriate?<br>5. Who should be allowed to access it?<br>6. How should the agent behave?<br>7. How will the output be tested?<br>8. How will the agent be exposed to the user?<br>9. How will performance and security be monitored?</p><p>Once these questions become clear, many Oracle AI Agent Studio scenarios become easier to reason through.</p><p>And that is the real purpose of preparation: not simply to select the correct answer in an assessment, but to understand how to design an AI agent that can work reliably in a real Oracle Fusion enterprise environment.</p><p><strong>Sample Quetions: Oracle Fusion AI Agent Studio Certified Developer Professional — Rel 26–2</strong></p><p><strong>Question 1</strong></p><p><strong>Which component within an AI Agent is used to apply Oracle Fusion role-based security settings?</strong></p><p>A. Business Object</p><p>B. Agent</p><p>C. Tools</p><p>D. Agent Team</p><p><strong>Correct answer: </strong>Agent Team</p><p><strong>Question 2</strong></p><p><strong>During testing, a Recruiting AI Agent gives unclear answers when candidates ask about job application status. What is the best way to refine the system prompt?</strong></p><p>A. Add more information to the prompt without changing its structure.</p><p>B. Rewrite the prompt as concise, step-by-step instructions that clearly define the agent’s expected actions.</p><p>C. Use a single JSON instruction for all prompt improvements.</p><p><strong>Correct answer: </strong>Rewrite the prompt as concise, step-by-step instructions that clearly define the agent’s expected actions.</p><p><strong>Question 3</strong></p><p><strong>A custom administrator role must manage agent permissions and access across several Fusion Cloud applications. What is the most appropriate security setup?</strong></p><p>A. Assign a custom role directly to users and manually configure separate security policies for every Agent Team.</p><p>B. Add the custom role and enable Access Intelligent Agent Chat only.</p><p>C. Enable permission groups for the custom role and assign the appropriate duties for each product family.</p><p><strong>Correct answer: </strong>Enable permission groups for the custom role and assign the appropriate duties for each product family.</p><p><strong>Question 4</strong></p><p><strong>An Agent supports questions related to staffing, scheduling, and customer complaints. How should the Topics be organized?</strong></p><p>A. Create one large Topic that combines all three areas for simplicity.</p><p>B. Create three separate Topics — staffing, scheduling, and customer complaints — with their own instructions.</p><p>C. Use a Deep Link for each area instead of creating Topics.</p><p><strong>Correct answer: </strong>Create three separate Topics — staffing, scheduling, and customer complaints — with their own instructions.</p><p><strong>Question 5</strong></p><p><strong>A Solution Architect is reviewing Oracle AI Studio automated testing results before production. Which metric is mainly used to determine whether an Agent’s responses are correct?</strong></p><p>A. User satisfaction scores from live A/B testing.</p><p>B. Correctness of the generated response compared with the expected answer.</p><p>C. Token efficiency relative to the test set average.</p><p>D. Response latency from input to output.</p><p><strong>Correct answer: </strong>Correctness of the generated response compared with the expected answer.</p><p><strong>Question 6</strong></p><p><strong>A customer service team wants to validate whether an Agent’s response format and wording are appropriate for representative use. What is the best testing approach?</strong></p><p>A. Compare the responses with FAQ documents and check structural consistency.</p><p>B. Have a small leadership group decide whether the answers sound clear.</p><p>C. Create realistic business scenarios, test the Agent’s responses, and compare them with the official company policy.</p><p>D. Collect feedback only after the Agent is released to production.</p><p><strong>Correct answer: </strong>Create realistic business scenarios, test the Agent’s responses, and compare them with the official company policy.</p><p><strong>Question 7</strong></p><p><strong>Which Enterprise Scheduler Service (ESS) process must run before an Agent can use a document as a retrieval source?</strong></p><p>A. HCM Analytics Initial Ingestion</p><p>B. Process Agent Documents</p><p>C. Mass Ingest Journey Data</p><p>D. Create Summaries Using Generative AI</p><p><strong>Correct answer: </strong>Process Agent Documents</p><p><strong>Question 8</strong></p><p><strong>A company wants users to interact with published Agents while restricting deployment and administration rights. Which two roles should be assigned?</strong></p><p>A. A custom user role that includes Access Intelligent Agent Chat privilege.</p><p>B. Application Implementation Consultant with the required permission groups and Access Intelligent Agent Chat privilege.</p><p>C. Application Implementation Consultant with permission groups for administrators plus a custom Explorer role.</p><p><strong>Correct answer: </strong>A custom user role that includes Access Intelligent Agent Chat privilege.; Application Implementation Consultant with the required permission groups and Access Intelligent Agent Chat privilege.</p><p><strong>Question 9</strong></p><p><strong>Which set of HTTP operations is supported when creating a business object for Oracle Fusion AI Agent Studio?</strong></p><p>A. HTTP TRACE, HTTP CONNECT, HTTP SEARCH, HTTP POST</p><p>B. TCP PATCH, HTTP HEAD, HTTP OPTIONS, HTTP GET</p><p>C. HTTP GET, HTTP POST, HTTP PUT, HTTP DELETE</p><p><strong>Correct answer: </strong>HTTP GET, HTTP POST, HTTP PUT, HTTP DELETE</p><p><strong>Question 10</strong></p><p><strong>A Benefits Agent produces confusing responses when users ask for current enrollment information. What prompt refinement is most appropriate?</strong></p><p>A. Add more policy terminology so the wording closely matches Benefits policies.</p><p>B. Include additional coverage and enrollment attributes to make the prompt longer.</p><p>C. Rephrase the instruction with a clear example such as, “Show my current health plan and what is covered.”</p><p><strong>Correct answer: </strong>Rephrase the instruction with a clear example such as, “Show my current health plan and what is covered.”</p><p><strong>Question 11</strong></p><p><strong>An HR organization wants escalation and response style to remain consistent across different teams. What is the best design?</strong></p><p>A. Configure one Business Object Tool that handles escalation and response styles.</p><p>B. Upload escalation and response-style documents and reference them from every Agent Team.</p><p>C. Create separate Topics for escalation logic and response style, then assign those Topics to the required Agent Teams.</p><p><strong>Correct answer: </strong>Create separate Topics for escalation logic and response style, then assign those Topics to the required Agent Teams.</p><p><strong>Question 12</strong></p><p><strong>What is the key difference between offline testing and A/B testing in Oracle AI Studio?</strong></p><p>A. Offline testing is fully automated, while A/B testing always requires manual intervention.</p><p>B. Offline testing evaluates responses against reference answers, while A/B testing can evaluate user satisfaction and effectiveness across configurations.</p><p>C. Offline testing focuses on latency, while A/B testing only measures engagement.</p><p>D. Offline testing optimizes performance, while A/B testing evaluates only configuration effectiveness.</p><p><strong>Correct answer: </strong>Offline testing evaluates responses against reference answers, while A/B testing can evaluate user satisfaction and effectiveness across configurations.</p><p><strong>Question 13</strong></p><p><strong>An administrator uploads new policy documents and needs them available to an Agent. What should be done?</strong></p><p>A. Link the policy documents to a Topic and attach that Topic to the Agent Team.</p><p>B. Upload the documents to the Document Tool and then run the Process Agent Documents ESS process.</p><p>C. Remove existing Agent Team tools and create a new Document Tool for each policy update.</p><p><strong>Correct answer: </strong>Upload the documents to the Document Tool and then run the Process Agent Documents ESS process.</p><p><strong>Question 14</strong></p><p><strong>An Oracle Fusion AI Agent accesses payroll information through an external REST API. What setup best supports secure access and logging?</strong></p><p>A. Allow all authenticated users to access payroll data and use HTTP endpoints for better performance.</p><p>B. Enable detailed audit logging for payroll access and use HTTPS endpoints.</p><p>C. Configure the REST Tool to log data-access events and restrict access to users with the designated HR role.</p><p>D. Require multifactor authentication for Agent access and let department managers approve all payroll queries.</p><p><strong>Correct answer: </strong>Configure the REST Tool to log data-access events and restrict access to users with the designated HR role.</p><p><strong>Question 15</strong></p><p><strong>An Agent needs to call an external service. Which information must be configured in the External REST Tool?</strong></p><p>A. The API response schema, supported error codes, and required mapping to Fusion objects.</p><p>B. The endpoint URL, HTTP operation, and authentication details for the external API.</p><p>C. A link to an external resource and a description of the available API tools.</p><p>D. A predefined set of answers and a short description of the Agent’s purpose.</p><p><strong>Correct answer: </strong>The endpoint URL, HTTP operation, and authentication details for the external API.</p><p><strong>Question 16</strong></p><p><strong>When evaluating Agent performance, latency and token usage are measured along with correctness. What does this help assess?</strong></p><p>A. Latency and token usage directly determine whether an answer is factually correct.</p><p>B. Latency and token usage measure efficiency and cost-performance, while correctness measures semantic alignment.</p><p>C. Latency and tokens confirm formatting compliance, while correctness measures only logical consistency.</p><p><strong>Correct answer: </strong>Latency and token usage measure efficiency and cost-performance, while correctness measures semantic alignment.</p><p><strong>Question 17</strong></p><p><strong>You are testing an Agent that gives product-representative answers. How can you confirm that the answers match official policy and business requirements?</strong></p><p>A. Ask customer service representatives to approve each answer individually before launch.</p><p>B. Test the Agent with realistic business questions and verify the answers against official policy.</p><p>C. Use only development-team documentation as the reference source.</p><p><strong>Correct answer: </strong>Test the Agent with realistic business questions and verify the answers against official policy.</p><p><strong>Question 18</strong></p><p><strong>A Guided Journey has been activated for a new Agent. What should be done in Visual Builder Studio to make sure it is visible on the homepage?</strong></p><p>A. Edit the Visual Builder Studio template and publish the homepage.</p><p>B. Enter the Guided Journey code in the homepage properties and use Preview to confirm the appearance.</p><p>C. Add the Guided Journey code only to the navigation menu.</p><p><strong>Correct answer: </strong>Enter the Guided Journey code in the homepage properties and use Preview to confirm the appearance.</p><p><strong>Question 19</strong></p><p><strong>A Finance AI Agent produces inconsistent prompt results because instructions are mixed together. What is the best prompt design approach?</strong></p><p>A. Put all prompt information into one long paragraph.</p><p>B. Combine persona, instructions, inputs, and outcomes into one narrative without sections.</p><p>C. Divide the prompt into clearly labeled sections such as persona, instructions, inputs, and desired outputs.</p><p><strong>Correct answer: </strong>Divide the prompt into clearly labeled sections such as persona, instructions, inputs, and desired outputs.</p><p><strong>Question 20</strong></p><p><strong>A Procurement Agent is consumed by downstream automation that expects JSON. How should the system prompt be updated?</strong></p><p>A. Keep the output format flexible because the content itself matters most.</p><p>B. Change the persona to plain text and let the downstream system handle formatting.</p><p>C. Explicitly require valid JSON output and prohibit Markdown formatting or backticks.</p><p><strong>Correct answer: </strong>Explicitly require valid JSON output and prohibit Markdown formatting or backticks.</p><p><strong>Question 21</strong></p><p><strong>A Performance Agent must use the same overall guidance for managers and employees but respond differently to each group. What is the best approach?</strong></p><p>A. Add a generic ‘role’ parameter to the context so the model customizes everything automatically.</p><p>B. Create distinct prompts or responses with targeted instructions for managers and employees.</p><p>C. Upload the company policy to a Document Tool and tell the Agent to cite it in every answer.</p><p><strong>Correct answer: </strong>Create distinct prompts or responses with targeted instructions for managers and employees.</p><p><strong>Question 22</strong></p><p><strong>A Guided Journey was created for a new Agent, but users cannot see the embedded Agent. Which two checks are appropriate?</strong></p><p>A. Verify that the Guided Journey is not in draft status.</p><p>B. Verify that the Guided Journey code was entered in the homepage-related property field.</p><p>C. Assume homepage layout customization after embedding will always override the Agent display settings.</p><p><strong>Correct answer: </strong>Verify that the Guided Journey is not in draft status.; Verify that the Guided Journey code was entered in the homepage-related property field.</p><p><strong>Question 23</strong></p><p><strong>You are configuring an AI Agent for HR questions and want it available from the employee homepage. Which two implementation steps are required?</strong></p><p>A. In Guided Journeys, create a new journey and add a task referencing the Agent before activating it and noting its generated code.</p><p>B. In Visual Builder Studio, edit page properties and enter the Guided Journey code in the dedicated homepage integration field.</p><p>C. In the Oracle Fusion Applications menu, duplicate the Agent Team and map it to the Guided Journey using a scheduled process.</p><p><strong>Correct answer: </strong>In Guided Journeys, create a new journey and add a task referencing the Agent before activating it and noting its generated code.; In Visual Builder Studio, edit page properties and enter the Guided Journey code in the dedicated homepage integration field.</p><p><strong>Question 24</strong></p><p><strong>Which duty role should be assigned to allow a user to interact with an AI Agent in the application?</strong></p><p>A. FAI GenAI Agent HCM Administrator Duty</p><p>B. FAI GenAI Manager Job</p><p>C. FAI GenAI Agent Runtime Duty</p><p>D. FAI GenAI Manager Duty</p><p><strong>Correct answer: </strong>FAI GenAI Agent Runtime Duty</p><p><strong>Question 25</strong></p><p><strong>A developer wants users to open the Payables Invoice work area directly from an Agent to edit invoices. Which configuration is appropriate?</strong></p><p>A. Add a Document Tool containing the invoice policy and page link, then assign it to the Agent Team.</p><p>B. Create a Deep Link using the instance-relative URL, associate it with a Deep Link Tool, and assign that tool to the Agent Team.</p><p>C. Configure an External REST Tool for the instance, specify the invoice endpoint, and assign it to the Agent Team.</p><p><strong>Correct answer: </strong>Create a Deep Link using the instance-relative URL, associate it with a Deep Link Tool, and assign that tool to the Agent Team.</p><p><strong>Question 26</strong></p><p><strong>An Agent integrates with an external API that requires OAuth 2.0 and specific data-encryption controls. Which option best meets the requirement?</strong></p><p>A. Basic authentication with SSL/TLS encryption.</p><p>B. OpenID Connect with RSA-based encryption.</p><p>C. OAuth 2.0 authentication with AES-256 data encryption.</p><p>D. OAuth 2.0 authentication with PGP encryption.</p><p><strong>Correct answer: </strong>OAuth 2.0 authentication with AES-256 data encryption.</p><p><strong>Question 27</strong></p><p><strong>A business unit needs only HR specialists to access a specific Agent Team. What is the correct access configuration?</strong></p><p>A. Assign the Agent Team to a custom role with Access Intelligent Agent Chat and restrict assignment to HR specialists.</p><p>B. Assign the Agent Team to the Application Implementation Consultant role and restrict access through menu configuration.</p><p>C. Assign the Agent Team to a custom HR specialist role with permission groups enabled and the relevant functional security policies.</p><p><strong>Correct answer: </strong>Assign the Agent Team to a custom HR specialist role with permission groups enabled and the relevant functional security policies.</p><p><strong>Question 28</strong></p><p><strong>How does Context information improve Agent Tracing?</strong></p><p>A. It automatically calculates testing correctness metrics for A/B outcomes.</p><p>B. It summarizes performance metrics, user satisfaction, and aggregate results.</p><p>C. It provides values such as usernames, start times, Agent memory, prompts, Topics, and instructions associated with each trace.</p><p><strong>Correct answer: </strong>It provides values such as usernames, start times, Agent memory, prompts, Topics, and instructions associated with each trace.</p><p><strong>Question 29</strong></p><p><strong>In Oracle AI Studio, how do Agent Tracing and Prompt Tracing differ?</strong></p><p>A. Agent Tracing provides step-by-step visualization of Agent execution, while Prompt Tracing focuses on the details of an individual prompt execution.</p><p>B. Agent Tracing focuses on latency and token usage, while Prompt Tracing analyzes model performance at the Agent level.</p><p>C. Agent Tracing is primarily for compliance workflows, while Prompt Tracing is dedicated to evaluation datasets and scoring.</p><p><strong>Correct answer: </strong>Agent Tracing provides step-by-step visualization of Agent execution, while Prompt Tracing focuses on the details of an individual prompt execution.</p><p><strong>Question 30</strong></p><p><strong>During offline testing, one Agent run has lower latency but slightly lower correctness than another run that is slower but more accurate. What is the best selection approach?</strong></p><p>A. Always choose the run with the lowest latency.</p><p>B. Select the run using comparative analysis, while treating correctness as the primary factor unless the SLA explicitly prioritizes cost or latency.</p><p>C. Merge both runs into a hybrid evaluation by averaging correctness and latency scores.</p><p><strong>Correct answer: </strong>Select the run using comparative analysis, while treating correctness as the primary factor unless the SLA explicitly prioritizes cost or latency.</p><p><strong>Question 31</strong></p><p><strong>A Benefits Enrollment Agent collects dependent information and later needs that information when the user selects a health plan. How should this be handled?</strong></p><p>A. Configure session variables to store dependent details so they remain available during the selected interaction.</p><p>B. Create separate workflows for dependents and health plans and treat them as unrelated steps.</p><p>C. Change the prompt so the model relies only on earlier conversation text rather than storing details explicitly.</p><p><strong>Correct answer: </strong>Configure session variables to store dependent details so they remain available during the selected interaction.</p><p><strong>Question 32</strong></p><p><strong>An Agent will access sensitive employee data and external APIs. How should the solution protect against unauthorized access, prompt injection, and data exposure?</strong></p><p>A. Grant broad access to all APIs and rely on security checks and monitoring after exchange.</p><p>B. Apply role-based access controls, enforce API authentication, and use Oracle guardrails for risky and prompt-injection inputs.</p><p>C. Use context-specific prompts and rely on external providers for all security enforcement.</p><p><strong>Correct answer: </strong>Apply role-based access controls, enforce API authentication, and use Oracle guardrails for risky and prompt-injection inputs.</p><p><strong>Question 33</strong></p><p><strong>An Agent must retrieve live inventory data from Fusion SCM, apply role-based access, and also support an external logistics integration. What design is appropriate?</strong></p><p>A. Combine a Document Tool for inventory records with a Deep Link Tool for logistics portals.</p><p>B. Add a Business Object Tool for inventory with role-based access and configure an external REST call for the required integration.</p><p>C. Use a Deep Link Tool to open the inventory dashboard and a Calculator Tool for stock estimates.</p><p><strong>Correct answer: </strong>Add a Business Object Tool for inventory with role-based access and configure an external REST call for the required integration.</p><p><strong>Question 34</strong></p><p><strong>After a quarterly update, an Agent can no longer retrieve invoice information. What is the most likely cause?</strong></p><p>A. The Agent’s prompt and Topics were overwritten during the update.</p><p>B. The quarterly update changed the API version path for custom business objects.</p><p>C. The update deprecated the REST API and all integrations must now use SOAP or a custom object.</p><p><strong>Correct answer: </strong>The quarterly update changed the API version path for custom business objects.</p><p><strong>Question 35</strong></p><p><strong>A financial services organization uses multiple Agents for regulatory compliance. Which design best maintains consistent guidance across the Agents?</strong></p><p>A. Add a system prompt to each Agent that dynamically generates the latest compliance instructions at runtime.</p><p>B. Create a centralized compliance Topic containing the required guidance and link that Topic to every Agent that needs it.</p><p>C. Upload the compliance manual into a Document Tool and require every Agent to search the manual.</p><p><strong>Correct answer: </strong>Create a centralized compliance Topic containing the required guidance and link that Topic to every Agent that needs it.</p><p><strong>Question 36</strong></p><p><strong>An HR Agent answers questions about payroll, leave, and benefits. How should the Agent’s Topics be structured?</strong></p><p>A. Create one broad Topic covering payroll, leave, and benefits together.</p><p>B. Use Deep Links so employees can select the HR area through navigation links.</p><p>C. Create separate Topics for payroll, leave, and benefits, and assign them to the Agent.</p><p>D. Upload HR policy documents and require the Agent to retrieve answers for all topics from them.</p><p><strong>Correct answer: </strong>Create separate Topics for payroll, leave, and benefits, and assign them to the Agent.</p><p><strong>Question 37</strong></p><p><strong>After configuring all tasks in a Guided Journey, what best describes the activation and preview requirements?</strong></p><p>A. Preview is mandatory, while activation is optional.</p><p>B. Activation and preview are both required.</p><p>C. Activation is required; preview is recommended.</p><p><strong>Correct answer: </strong>Activation is required; preview is recommended.</p><p><strong>Question 38</strong></p><p><strong>A global company wants customer-support Agents to respond in several languages and according to regional policies. What is the best Topic design?</strong></p><p>A. Upload regional policy documents into a Document Tool.</p><p>B. Use one large Topic containing all languages and regional policies.</p><p>C. Create Topics based on language and region.</p><p><strong>Correct answer: </strong>Create Topics based on language and region.</p><p><strong>Question 39</strong></p><p><strong>You are building an Agent to help employees retrieve pay records, view payslip details, and calculate tax. Users should see only information they are authorized to access. Which setup is appropriate?</strong></p><p>A. Use a Business Object Tool for pay data, a Deep Link Tool for payslip details, and a Context Tool to enforce user-level access controls.</p><p>B. Use an External REST Tool for an external payroll provider and a Document Tool for pay statements.</p><p>C. Configure a Document Tool with a policy PDF and assign a Topic for calculations.</p><p>D. Use a Calculator Tool for tax calculations and a Document Tool for pay-policy information.</p><p><strong>Correct answer: </strong>Use a Business Object Tool for pay data, a Deep Link Tool for payslip details, and a Context Tool to enforce user-level access controls.</p><p><strong>Question 40</strong></p><p><strong>A financial services company needs detailed regulatory, tax, and audit guidance and wants clear separation between the areas. How should the Topics be configured?</strong></p><p>A. Use one Business Object Tool for regulatory, tax, and audit queries.</p><p>B. Create a single Topic containing regulatory, tax, and audit guidance.</p><p>C. Create separate Topics for regulatory, tax, and audit.</p><p><strong>Correct answer: </strong>Create separate Topics for regulatory, tax, and audit.</p><p><strong>Question 41</strong></p><p><strong>A Leave and Absence Agent often receives vague requests such as ‘show leave’ and responds inconsistently. What should be added to improve accuracy?</strong></p><p>A. Define clear rules in the system prompt for handling ambiguous queries and instruct the Agent to ask for clarification when necessary.</p><p>B. Keep the prompt minimal and let the model infer intent entirely from user context.</p><p>C. Require users to phrase all questions as complete sentences.</p><p><strong>Correct answer: </strong>Define clear rules in the system prompt for handling ambiguous queries and instruct the Agent to ask for clarification when necessary.</p><p><strong>Question 42</strong></p><p><strong>An Oracle Fusion implementation team needs separate administrator access for both HCM and SCM, while permissions must remain managed according to product areas. Which setup is appropriate?</strong></p><p>A. Enable permission groups and assign the Supply Chain Application Administrator for SCM; create a custom admin role for HCM with the required duties.</p><p>B. Assign only the Application Implementation Consultant role to every administrator for all setup needs.</p><p>C. Assign the HCM Application Administrator and SCM Application Administrator roles without enabling permission groups.</p><p><strong>Correct answer: </strong>Enable permission groups and assign the Supply Chain Application Administrator for SCM; create a custom admin role for HCM with the required duties.</p><p><strong>Question 43</strong></p><p><strong>A manufacturing company is configuring an Agent for order management and wants to read and update Fusion transactional records while using the signed-in employee’s privileges. Which components should be included?</strong></p><p>A. Deep Link Tool</p><p>B. Document Tool</p><p>C. External REST Tool</p><p>D. Business Object Tool</p><p>E. User Session Tool</p><p><strong>Correct answer: </strong>Deep Link Tool; Business Object Tool; User Session Tool</p><p><strong>Question 44</strong></p><p><strong>A developer is creating a Business Object Tool to query and update Benefits Enrollment data through a REST API. What should be configured for the enrollment result identifier?</strong></p><p>A. Add the Enrollment Result ID as a parameter in the Parameters and Business Object fields.</p><p>B. Add the enrollment result identifier as a parameter directly in the resource path, such as /benefitEnrollments/{Enrollment ResultId}.</p><p>C. Use an Add from Specification option to automatically create a Business Object for benefit enrollments.</p><p><strong>Correct answer: </strong>Add the enrollment result identifier as a parameter directly in the resource path, such as /benefitEnrollments/{Enrollment ResultId}.</p><p><strong>Question 45</strong></p><p><strong>You are deploying a Guided Journey that contains an Agent in Oracle Fusion Applications. During deployment, where should you configure the required setting so users can access the Agent journey code?</strong></p><p>A. Modify navigation menu options in Visual Builder Studio.</p><p>B. Update the Page Properties section in Visual Builder Studio to include the Guided Journey code.</p><p>C. Create a new business object in Visual Builder Studio to store the Agent journey code.</p><p><strong>Correct answer: </strong>Update the Page Properties section in Visual Builder Studio to include the Guided Journey code.</p><p><strong>Question 46</strong></p><p><strong>Which capability is provided by the pre-built ‘My Team’ Business Object Tool in Oracle AI Agent Studio?</strong></p><p>A. It uses an employee assignment function to retrieve team details.</p><p>B. It provides a pre-built function for searching workers to query team members.</p><p>C. It retrieves only workers who report directly to the logged-in user.</p><p><strong>Correct answer: </strong>It provides a pre-built function for searching workers to query team members.</p><p><strong>Question 47</strong></p><p><strong>A Benefits Policy Agent in Debug mode is slow because it repeatedly calls the Document Tool, retrieves too many chunks, and uses an overly verbose prompt. Which actions should be taken?</strong></p><p>A. Review the Debug trace for redundant tool calls and reduce the document set or adjust chunking to minimize retrieval overhead.</p><p>B. Switch to a larger LLM because larger models always respond faster.</p><p>C. Increase the number of retrieved chunks so the model receives more context for every question.</p><p>D. Review and streamline the prompt and Topic instructions to reduce verbosity and overall request size, then retest latency.</p><p><strong>Correct answer: </strong>Review the Debug trace for redundant tool calls and reduce the document set or adjust chunking to minimize retrieval overhead.; Review and streamline the prompt and Topic instructions to reduce verbosity and overall request size, then retest latency.</p><p><strong>Question 48</strong></p><p><strong>An Outbound Compliance Agent retrieves a user’s internal policy documents and an external REST resource, but Debug mode shows repeated tool calls and inconsistent guidance based on internal policy. What change should be made?</strong></p><p>A. Direct the Agent to use the Document Tool for compliance and keep the Deep Link Tool for navigation.</p><p>B. Replace the document-based policy reference with a Business Object configuration so the Agent retrieves similar guidance from transaction endpoints.</p><p>C. Disable the Deep Link capability so users cannot navigate away from the Agent.</p><p>D. Keep all existing Topics active and upload a larger internal PDF so the Agent has more information before responding.</p><p><strong>Correct answer: </strong>Direct the Agent to use the Document Tool for compliance and keep the Deep Link Tool for navigation.</p><p><strong>Question 49</strong></p><p><strong>A multinational company is launching an Agent to automate leave requests and provide real-time status updates. The Agent must access sensitive employee data through Business Object Tools and call external REST APIs. What is the best security approach?</strong></p><p>A. Grant unrestricted access roles and enable broad tool access for operational flexibility.</p><p>B. Configure Agent access using least-privilege roles and internal APIs, apply role-based and Agent-level restrictions, and require token keys for external API authentication.</p><p>C. Limit Agent Tools to required business objects and APIs but rely only on user roles for external API requests.</p><p>D. Deploy the Agent in a separate test environment and rely on extended debugging and logging as the primary security mechanism.</p><p><strong>Correct answer: </strong>Configure Agent access using least-privilege roles and internal APIs, apply role-based and Agent-level restrictions, and require token keys for external API authentication.</p><p>#OracleFusion #OracleAI #AIAgents #AgenticAI #GenerativeAI #OracleCertification #OracleCloud #PromptEngineering #EnterpriseAI #OracleAIAgentStudio</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=47aada95f536" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[AI Tool Integration with Oracle Integration Cloud (OIC)]]></title>
            <link>https://medium.com/@jhasamiree1/ai-tool-integration-with-oracle-integration-cloud-oic-30648da2b8e1?source=rss-8a00b14ded6e------2</link>
            <guid isPermaLink="false">https://medium.com/p/30648da2b8e1</guid>
            <dc:creator><![CDATA[Samir Jha|Oracle Architect (OCI|OIC|SaaS Appl)]]></dc:creator>
            <pubDate>Sun, 26 Jul 2026 09:51:36 GMT</pubDate>
            <atom:updated>2026-07-26T09:51:36.681Z</atom:updated>
            <content:encoded><![CDATA[<h3>Introduction</h3><p>This blog demonstrates how to build an <strong>AI Tool</strong> embedded within an <strong>Oracle Integration Cloud (OIC) Gen3 Project</strong> using the built-in AI capabilities of OIC. The AI tool is designed to retrieve the corresponding village or post office name for a given <strong>PIN Code</strong> by invoking an external REST API.</p><p>The primary focus of this blog is not only on building the AI tool but also on understanding <strong>how to configure and use AI Tools in OIC Gen3</strong>. Since this is a relatively new feature, many OIC developers may not yet be familiar with its capabilities and implementation. This step-by-step guide aims to provide a practical introduction, helping you create, configure, and effectively use AI Tools within your OIC Gen3 projects.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/906/1*JOgbCBUzj_MBDIOc-JW5rQ.png" /><figcaption>WorkFlow</figcaption></figure><h3>Step 1: Create the Integration</h3><p>Log in to <strong>Oracle Integration Cloud (OIC) Gen3</strong> and navigate to <strong>Projects</strong>. The left-hand navigation pane displays the various OIC Gen3 capabilities, including <strong>Integration</strong>, <strong>AI Agents</strong>, <strong>Robot</strong>, <strong>Human in the Loop</strong>, <strong>Decision</strong>, <strong>Healthcare</strong>, <strong>B2B</strong>, and <strong>Knowledge Base</strong>.</p><p><em>Image: Projects — OIC Gen3</em></p><p>Before creating the AI Tool, you first need to build a <strong>standard Integration</strong> that retrieves post office details for a given PIN Code. This integration will act as the backend service that the AI Tool invokes.</p><p>For this example, I have used the free Postal PIN Code REST API provided by <a href="https://api.postalpincode.in/"><strong>https://api.postalpincode.in</strong></a>, which returns post office information based on the supplied PIN Code. The AI Tool will later invoke this integration to retrieve and process the required data.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/60/1*k3B8Cg6Gt3uoe0cW_kgfEQ.png" /><figcaption>Image from Projects- OIC Gen3</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*cA2bqaI_sCk2xJzVFbOvIw.png" /><figcaption>pinCode OIC</figcaption></figure><h3>Step 2: Create an AI Tool</h3><p>Once the integration created in <strong>Step 1</strong> has been deployed, navigate to the <strong>AI Agents</strong> section from the left-hand menu in your <strong>OIC Gen3 Project</strong>.</p><p>Within <strong>AI Agents</strong>, you will find the following four sections:</p><ul><li><strong>Agents</strong> — Create and manage AI agents.</li><li><strong>Agent Patterns</strong> — Define reusable agent orchestration patterns.</li><li><strong>Tools</strong> — Create and manage tools that agents can invoke.</li><li><strong>Prompt Templates</strong> — Define reusable prompts and instructions for AI agents.</li></ul><p>Select <strong>Tools</strong> and click <strong>Create Tool</strong>. When prompted for the tool type, choose <strong>OIC Integration</strong>, and then select the integration that you created in <strong>Step 1</strong>.</p><p>This creates an AI Tool that acts as a wrapper around your OIC integration, allowing AI agents to invoke it to retrieve post office details for a specified PIN Code.</p><p>In the next steps, we’ll configure the tool’s input and output parameters, along with the instructions that guide the AI on when and how to invoke the tool.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*fqDi0vEqGbPpjIt8Fjcj_A.png" /><figcaption>create Tool</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*D8oTI11aSANrk_-waC7zCQ.png" /><figcaption>PIN_CODE_POC Tool</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*fbi3fS3PnL7Q4ZAp6hNoHQ.png" /><figcaption>Description and Guidelines</figcaption></figure><h3>Step 3: Create an Agent Pattern</h3><p>With the integration and AI Tool in place, the next step is to create an <strong>Agent Pattern</strong>. Navigate to <strong>AI Agents → Agent Patterns</strong> and click <strong>Create Agent Pattern</strong>.</p><p>During the creation process, you will be prompted to select a <strong>Pattern Recipe</strong>. Oracle provides multiple pattern recipes, such as <strong>ReAct</strong>, <strong>ReAct for OCI GenAI</strong>, <strong>Plan GenAI</strong>, and <strong>Plan</strong>, each designed for different AI interaction models. For this example, we will use the <strong>ReAct</strong> pattern, as it is ideal for scenarios where the AI needs to reason about the user’s request, invoke tools, process the returned information, and generate an accurate response.</p><p>After selecting the pattern recipe, the next and most important step is to define the <strong>Guidelines</strong> for the agent. These instructions tell the AI <strong>when to invoke the tool, what inputs are mandatory, how to process the tool response, and what output should be returned</strong>. Well-defined guidelines ensure that the AI consistently uses the tool instead of relying on its own knowledge, resulting in accurate and predictable responses.</p><p>Here I have created the connection using <a href="https://platform.openai.com/">https://platform.openai.com/</a></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*FsT2y6RAJj52ln3_2yHwlg.png" /><figcaption>Agent patterns</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*9KecY1G9YudqYa18t-8Ozw.png" /><figcaption>Guidelines in Pattern</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/619/1*GRHN66eL7-ZXWXx24WDoXg.png" /><figcaption>llm Connection</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/375/1*QTIxZS6XnFsusFoNYRPihg.png" /></figure><h3>Choosing the Right Pattern Recipe</h3><p>When creating an AI Agent in OIC Gen3, you need to select a <strong>Pattern Recipe</strong>. A pattern defines how the AI agent thinks, plans, and interacts with tools to complete a task. Oracle provides multiple patterns, each designed for different use cases.</p><h3>ReAct (26.4.1)</h3><p>The <strong>ReAct (Reason + Act)</strong> pattern enables the AI agent to reason through a user’s request, invoke the required tools, analyze the results, and continue this cycle until it produces the final response. This is the recommended choice for most tool-driven AI scenarios.</p><h3>ReAct for OCI GenAI (26.4.4)</h3><p>This pattern is an optimized version of ReAct specifically designed for <strong>Oracle Cloud Infrastructure (OCI) Generative AI</strong> models. It provides enhanced integration with OCI GenAI services while following the same Reason + Act approach.</p><h3>Plan GenAI (26.4.1)</h3><p>The <strong>Plan GenAI</strong> pattern uses Generative AI to first create an execution plan and then carries out the planned steps using the available tools. It is well suited for complex workflows that require multiple coordinated actions.</p><h3>Plan (26.4)</h3><p>The <strong>Plan</strong> pattern follows a structured planning approach by breaking a request into sequential steps before executing them. It is ideal for predictable business processes where tasks must follow a predefined order.</p><p>Now, teh moment the Agent pattern is saved, it creates a ReAct OIC is generated</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/852/1*v9UIQk0Glt1GgtcnOE4M4w.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*zR2R134xYFcDwhbXUp4jNg.png" /><figcaption>ReAct OIC</figcaption></figure><h3>Step 5: Create the AI Agent</h3><p>The final step is to create the <strong>AI Agent</strong>. Navigate to <strong>AI Agents → Agents</strong> and click <strong>Create Agent</strong>.</p><p>While creating the agent, provide a meaningful <strong>Name</strong> and <strong>Description</strong>, and then select the <strong>Agent Pattern</strong> created in <strong>Step 3</strong>. The agent will inherit the pattern recipe and guidelines defined in the selected Agent Pattern.</p><p>Next, review and update the <strong>Guidelines</strong> if required. These instructions define the agent’s behavior, including when to invoke the tool, how to process the response, and how to generate the final output.</p><p>Once the configuration is complete, save and deploy the agent. Your AI Agent is now ready to accept user requests, invoke the OIC Integration through the configured AI Tool, and return accurate responses based on the data retrieved from the external API.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*6XI1dOLgLe06YLcWPZ7iMA.png" /><figcaption>AI Agent</figcaption></figure><h3>Step 6: Test the AI Agent</h3><p>Now it’s time to test the AI Agent.</p><p>The Postal PIN Code API returns <strong>all post offices associated with the specified PIN Code</strong>. However, in this example, our requirement is to retrieve <strong>only the details of the specific village</strong> provided by the user.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*eDCzidD-p8shx2VTMgLcBw.png" /><figcaption>Results..</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/897/1*hD5ymKXqpnVCJdGmcdYLIg.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*BpPeqB5AQOaviPcxJR89pg.png" /></figure><p>Token Usage Token :</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*_thNAjGbTXxmYeRny50B4w.png" /></figure><p>#Oracle #OracleCloud #OracleIntegrationCloud #OIC #OICGen3 #OracleIntegration #OCI #OracleAI #OracleAIAgentStudio #AIAgents #AITools #AgenticAI #GenerativeAI #EnterpriseAI #RESTAPI #APIIntegration #CloudIntegration #PromptEngineering #ReActPattern #Automation #OracleCommunity #Developer #MediumBlog</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=30648da2b8e1" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Oracle Fusion AI Studio — Monitoring and Evaluating AI Agents — Part IV]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/@jhasamiree1/monitoring-and-evaluating-ai-agents-in-oracle-fusion-ai-studio-1df30c7c91eb?source=rss-8a00b14ded6e------2"><img src="https://cdn-images-1.medium.com/max/814/1*PsRL1TmPf8PlKRKgHxle_Q.png" width="814"></a></p><p class="medium-feed-snippet">Introduction:</p><p class="medium-feed-link"><a href="https://medium.com/@jhasamiree1/monitoring-and-evaluating-ai-agents-in-oracle-fusion-ai-studio-1df30c7c91eb?source=rss-8a00b14ded6e------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://medium.com/@jhasamiree1/monitoring-and-evaluating-ai-agents-in-oracle-fusion-ai-studio-1df30c7c91eb?source=rss-8a00b14ded6e------2</link>
            <guid isPermaLink="false">https://medium.com/p/1df30c7c91eb</guid>
            <category><![CDATA[oracle-ai-agent-studio]]></category>
            <dc:creator><![CDATA[Samir Jha|Oracle Architect (OCI|OIC|SaaS Appl)]]></dc:creator>
            <pubDate>Tue, 14 Jul 2026 09:46:04 GMT</pubDate>
            <atom:updated>2026-09-24T10:02:55.352Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[Oracle Fusion AI Studio — Agentic AI Design Patterns: Workflow vs. Supervisor — Part III]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/@jhasamiree1/oracle-fusion-agentic-ai-design-patterns-workflow-vs-supervisor-39bac12a5d17?source=rss-8a00b14ded6e------2"><img src="https://cdn-images-1.medium.com/max/698/1*kQciIwOSd0WexWGMYKK2Xw.png" width="698"></a></p><p class="medium-feed-snippet">Introduction</p><p class="medium-feed-link"><a href="https://medium.com/@jhasamiree1/oracle-fusion-agentic-ai-design-patterns-workflow-vs-supervisor-39bac12a5d17?source=rss-8a00b14ded6e------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://medium.com/@jhasamiree1/oracle-fusion-agentic-ai-design-patterns-workflow-vs-supervisor-39bac12a5d17?source=rss-8a00b14ded6e------2</link>
            <guid isPermaLink="false">https://medium.com/p/39bac12a5d17</guid>
            <category><![CDATA[oracle-fusion]]></category>
            <dc:creator><![CDATA[Samir Jha|Oracle Architect (OCI|OIC|SaaS Appl)]]></dc:creator>
            <pubDate>Wed, 17 Jun 2026 06:26:46 GMT</pubDate>
            <atom:updated>2026-09-24T10:02:23.755Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[Oracle Fusion AI Studio — Agentic AI Workflow Learning — Part II]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/@jhasamiree1/oracle-agentic-ai-workflow-learning-671e3f42c963?source=rss-8a00b14ded6e------2"><img src="https://cdn-images-1.medium.com/max/658/1*9n4oc_yJWdVDpVqtcEFoIg.png" width="658"></a></p><p class="medium-feed-snippet">What Are Workflow Agents?</p><p class="medium-feed-link"><a href="https://medium.com/@jhasamiree1/oracle-agentic-ai-workflow-learning-671e3f42c963?source=rss-8a00b14ded6e------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://medium.com/@jhasamiree1/oracle-agentic-ai-workflow-learning-671e3f42c963?source=rss-8a00b14ded6e------2</link>
            <guid isPermaLink="false">https://medium.com/p/671e3f42c963</guid>
            <category><![CDATA[oracle-fusion]]></category>
            <category><![CDATA[agentic-ai]]></category>
            <dc:creator><![CDATA[Samir Jha|Oracle Architect (OCI|OIC|SaaS Appl)]]></dc:creator>
            <pubDate>Fri, 24 Apr 2026 11:24:33 GMT</pubDate>
            <atom:updated>2026-09-24T10:01:28.680Z</atom:updated>
        </item>
    </channel>
</rss>