<?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 Jake Therianos on Medium]]></title>
        <description><![CDATA[Stories by Jake Therianos on Medium]]></description>
        <link>https://medium.com/@jakewlt5?source=rss-de2b44c0b55c------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*aoInu_JGDN_A7fhaZzfJNw.jpeg</url>
            <title>Stories by Jake Therianos on Medium</title>
            <link>https://medium.com/@jakewlt5?source=rss-de2b44c0b55c------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Fri, 20 Nov 2026 10:04:53 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@jakewlt5/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[AI-Ready Data: The Missing Link Between Context and Intelligence AI-ready data isn’t one layer.]]></title>
            <link>https://medium.com/@jakewlt5/context-categories-why-ai-agents-need-more-than-a-semantic-layer-a94d606fd918?source=rss-de2b44c0b55c------2</link>
            <guid isPermaLink="false">https://medium.com/p/a94d606fd918</guid>
            <category><![CDATA[semantic-layer]]></category>
            <category><![CDATA[context]]></category>
            <category><![CDATA[databricks]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[ontology]]></category>
            <dc:creator><![CDATA[Jake Therianos]]></dc:creator>
            <pubDate>Fri, 23 Oct 2026 13:40:21 GMT</pubDate>
            <atom:updated>2026-10-23T13:58:56.925Z</atom:updated>
            <content:encoded><![CDATA[<h3>AI-Ready Data: The Missing Link Between Context and Intelligence</h3><h4>AI-ready data isn’t one layer. It’s three. And most of the conversation is missing the third.</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*GIO6fN2XuQjxIvajKrBw2w.gif" /></figure><h3>What does AI-ready data really mean?</h3><p>When organizations say they want “AI-ready data,” they’re often talking about building a semantic layer or context layer so agents can understand their data. That’s important. But it isn’t the whole story.</p><p>I’m talking specifically about agents that need to discover, interpret, and retrieve enterprise information to answer questions or complete work: AI coworkers, analytics agents, research assistants, and agents embedded in business workflows. Not every agent fits that mold. A coding agent working primarily in a repository, for example, may depend more on source code, tests, and development tools than on enterprise data semantics.</p><p>At a high level, these agents need two things: a <em>map</em> and the <em>facts</em>. <strong>The map</strong> tells them where to look and how to interpret what they find; <strong>the facts</strong> are the actual records, documents, and relationships needed to answer the question. I split the map into two layers because finding the right data and understanding its business meaning are different problems — and solving one doesn’t automatically solve the other. Together with the facts, that gives us three practical layers of enterprise context.</p><h3>Why is context retrieval becoming more important than context stuffing?</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Me4bOXgqb7HteU_uTeEAXQ.gif" /></figure><p>The early pattern for many language-model applications was to assemble a large prompt: instructions, reference material, perhaps some retrieved passages, and the user’s question. But today’s agents can choose tools, inspect results, and fetch additional information as the task unfolds. The emphasis shifts from supplying all possible context up front to making the right context discoverable on demand.</p><p>Think about the context an agent already handles: instructions tell it how to behave; skills describe repeatable workflows; memory and conversation history preserve relevant prior information; and tool descriptions advertise what actions are available. These sources aren’t necessarily indexed in the same technical way. Skills may be selected through names and descriptions, tools through schemas, memories through relevance filtering, and documents through search indexes. But they have a common architectural idea: a way to locate the information without blindly dumping everything into the model’s context window.</p><p>Enterprise data takes that challenge a step further. An agent might need to find the authoritative revenue dataset among thousands of tables, understand the CFO’s definition of revenue, and then calculate last quarter’s actual number. That’s not one retrieval problem. It’s several different jobs chained together.</p><h3>What are the three layers of enterprise context?</h3><p>I find it useful to group enterprise data context by the job it performs for the agent. This is a conceptual model, not a strict taxonomy: a catalog page might contain both technical metadata and business rules, and an ontology can bridge semantics and domain entities. The boundaries get blurry quickly. That’s okay.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ZLmnIVDv69NfK7UZ9s3vqw.png" /></figure><h4>Layer 1: Data-estate context — where is the right information?</h4><p>Layer 1 describes the data estate: catalogs, tables, columns, ownership, lineage, access permissions, freshness, quality, and which assets are considered authoritative. It’s the map of what exists and where it lives.</p><p>Suppose an agent is asked, “What was US revenue for Product X last quarter?” Before generating SQL, it needs to find the right data among perhaps 20,000 tables. A fluent SQL query against the wrong table is still a wrong answer. Layer 1 helps the agent discover and choose the right source.</p><h4>Layer 2: Business semantic context — what does the information mean?</h4><p>Finding a table isn’t enough. Layer 2 defines the business concepts that govern how data should be interpreted: metrics, dimensions, synonyms, valid joins, calculation rules, canonical entities, and accepted query patterns.</p><p>For example, “revenue” might mean gross sales to one team and recognized net revenue to another. A model can understand both phrases in general, but it can’t know the company’s authoritative definition without context. Layer 2 can specify that net revenue equals gross sales minus rebates and returns, along with the correct fields and business calendar.</p><p>Layers 1 and 2 are largely what people mean when they discuss the enterprise semantic layer or context layer. They describe information about the data: how to locate it, whether to trust it, and how to use it.</p><h4>Layer 3: Domain data context — what are the actual facts?</h4><p>Layer 3 is the information itself: the sales transactions, customer records, support tickets, inventory events, document passages, or relationships between entities that answer the question. If Layer 2 defines how net revenue should be calculated, Layer 3 is the actual sales and rebate data from which an answer like “$14.2 million” is derived.</p><p>That might sound obvious, but it changes how we talk about agent architecture. An agent can retrieve Layer 3 with SQL, document search, APIs, or graph traversal. The semantic layer does not have to contain every business fact. It has to help the agent reach the right facts reliably.</p><h3>Does better AI eliminate the need for data modeling?</h3><p>Not remotely. In fact, this is one of the parts of the AI-ready-data conversation I think gets overlooked.</p><p>Data modeling is the work of defining how information is collected, organized, related, and stored so that it can be used consistently. BI teams have been doing this for years. A star schema organizes facts and dimensions for efficient analytics. A normalized relational model supports operational consistency. A document model preserves flexible records. An ontology defines domain concepts and allowed relationships; a knowledge graph can represent the actual entities and connections that instantiate those concepts.</p><p>These models solve different problems. A sales performance question may be best served by a fact table and dimensions. Finding the relevant paragraph of a policy may call for document retrieval. Tracing an unknown number of tiers through a supplier network may benefit from graph modeling and traversal. A knowledge graph, incidentally, does not automatically require a graph database: bounded entity relationships can often be stored in relational tables and queried with joins.</p><p>Better models don’t make messy data magically self-explanatory. If identifiers disagree across systems, records are duplicated, or business definitions are unresolved, an intelligent agent inherits those problems. Layer 1 and Layer 2 can guide retrieval, but substantial AI-readiness work still happens in the design, quality, and relationships of Layer 3 data.</p><p>My rule of thumb is to let the question dictate the model. Start with the simplest representation and retrieval pattern that answers it correctly. Add deeper relationship modeling when the use case genuinely needs it — not because “knowledge graph” is the term everyone is using this month.</p><h3>How does Databricks Genie retrieve all three layers?</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*BQOqcdNDjZj5gtz1BxQvJg.gif" /></figure><p>This is where Databricks Genie gets interesting to me — not just as a conversational analytics interface, but as an enterprise-context retrieval mechanism.</p><p>Genie Ontology is the governed, business-aware map. It combines curated Unity Catalog semantics — such as Metric Views, Domains, and UC Pages — with context inferred from existing assets and usage. When a question comes in, Genie can search and rank relevant context rather than requiring someone to hand-pick every definition and data asset in advance. That is the role of the map: help the agent determine which information matters and how to interpret it.</p><p>But Genie doesn’t stop at dynamically retrieving the map. It also uses that context to retrieve underlying facts, including querying structured data through SQL and working with supported unstructured sources. In the three-layer framing, it can use Layers 1 and 2 to navigate toward Layer 3. That distinction matters: getting the definition of revenue isn’t the same as retrieving and calculating the revenue number.</p><p>Imagine asking, “Which accounts saw declining usage, have open support issues, and need attention this week?” The agent must discover account, product-usage, and support sources; interpret what counts as declining usage and an open issue; retrieve the actual records; and combine them into a useful response. The compelling capability is not any single query. It’s coordinating discovery, interpretation, and fact retrieval within one workflow.</p><p>Genie One provides an AI coworker experience for business users, while the Genie One MCP server exposes that data-aware capability to external agents. In other words, the agent you already use can ask Genie for a grounded answer rather than attempting to reverse-engineer your company’s metrics from raw schemas.</p><h3>What Happens When Enterprise Data Meets Enterprise Knowledge?</h3><p>So far, we’ve focused on <strong>enterprise data retrieval</strong>: using business semantics to find and query the facts buried in your lakehouse. But much of an organization’s knowledge lives elsewhere, scattered across Google Drive, SharePoint, Slack, Confluence, Jira, and other applications.</p><p>This is where <strong>enterprise knowledge retrieval</strong> comes in. Think Glean-like enterprise search across organizational content, complemented by live MCP connectors for retrieving information directly from source systems.<a href="https://www.databricks.com/blog/how-genie-ontology-powers-product-development-databricks"> Genie One is already bringing these capabilities together</a>.</p><p>Now imagine an AI coworker querying the lakehouse to identify which products drove last quarter’s revenue decline, then searching internal documents and conversations to understand <em>why</em>.</p><p>The combination is what makes Genie One particularly interesting. It brings together<a href="https://docs.databricks.com/gcp/en/genie-one/external-sources"> enterprise search and connected tools</a> with governed analytics through a common agent experience.</p><p><strong>The bigger vision is unified enterprise context retrieval:</strong> giving AI agents a way to find the right information, regardless of where it lives.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*7-SUUyTxm0cb9C4vBZ1paQ.png" /></figure><h3>What should teams prioritize first?</h3><p>If you’re trying to make data useful to AI agents, I would focus on three things. First, make authoritative assets discoverable: document ownership, meaning, freshness, and permissions. Second, curate the business semantics that routinely cause ambiguity: metrics, joins, definitions, and rules. Third, invest in the underlying domain model and data quality so retrieval returns coherent facts rather than technically valid nonsense.</p><p>I would not begin by attempting to model the entire enterprise into one universal ontology or knowledge graph. Start with the questions people actually need answered, then deepen the context and modeling where accuracy demands it.</p><h3>Key takeaways</h3><ul><li>Layer 1 describes where the data is. Layer 2 explains what it means. Layer 3 contains the facts.</li><li>Genie’s agent harness is tightly integrated with its ontology for fast, intelligent retrieval across all three context layers, powering AI coworkers and custom agents.</li></ul><h3>Further reading</h3><p><a href="https://contextandchaos.substack.com/">Context and Chaos Substack</a> — Great source of content and opinions on this topic</p><p><a href="https://docs.databricks.com/aws/en/genie/genie-ontology">Genie Ontology</a></p><p><a href="https://docs.databricks.com/aws/en/agents/mcp-tools/genie-mcp">Genie One MCP server</a></p><p><a href="https://docs.databricks.com/gcp/en/genie-one/external-sources">Genie One external connections</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a94d606fd918" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>