<?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 Maud Parrot on Medium]]></title>
        <description><![CDATA[Stories by Maud Parrot on Medium]]></description>
        <link>https://medium.com/@maud.parrot_925?source=rss-c887c489dd6a------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*0xoYmr7y1lsBPkO2</url>
            <title>Stories by Maud Parrot on Medium</title>
            <link>https://medium.com/@maud.parrot_925?source=rss-c887c489dd6a------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Thu, 19 Nov 2026 08:20:30 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@maud.parrot_925/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[Our users knew exactly which documents they needed. Our assistant didn’t. Here’s what we built.]]></title>
            <link>https://medium.com/doctrine/our-users-knew-exactly-which-documents-they-needed-our-assistant-didnt-here-s-what-we-built-c4fbf3a7ed95?source=rss-c887c489dd6a------2</link>
            <guid isPermaLink="false">https://medium.com/p/c4fbf3a7ed95</guid>
            <category><![CDATA[engineering]]></category>
            <category><![CDATA[ux-design]]></category>
            <category><![CDATA[rags]]></category>
            <category><![CDATA[people]]></category>
            <category><![CDATA[chatbots]]></category>
            <dc:creator><![CDATA[Maud Parrot]]></dc:creator>
            <pubDate>Thu, 04 Jun 2026 11:56:22 GMT</pubDate>
            <atom:updated>2026-06-22T08:15:45.627Z</atom:updated>
            <content:encoded><![CDATA[<p>This article is based on a subject we dig into our tech days — a recurring ritual where we set aside time to explore ideas and pain points that fall outside the regular roadmap. As such, the solution presented here is exploratory and not intended for production as-is, though it’s something we may consider rolling out in the coming months. The project involved <a href="https://medium.com/@josselin.pennors">Josselin Pennors</a>, <a href="https://medium.com/@david.clochard77">David Clochard</a>, and <a href="https://medium.com/@maud.parrot_925">Maud Parrot</a>.</p><p>A lawyer opens the Assistant and types: <em>“Get inspired from </em><em>contract_template_3.pdf to fill in the contract for the Mrs. Blum case.&quot;, s</em>he knows exactly which document she needs. But our Assistant doesn&#39;t, and silently uses the wrong one. This is the problem we set out to fix.</p><h3>Context</h3><p>A key focus of our current development at Doctrine is the integration of clients’ internal knowledge. This is achieved by indexing external sources such as SharePoint and by allowing users to upload documents into dedicated internal knowledge bases. This condensed knowledge is then available in our Assistant, ready to be consumed to answer our client&#39;s needs.</p><p>But interviews with beta testers revealed something interesting: they often already know which document they need. Internally, we found that 13% of SharePoint queries come from users who have a specific document or folder in mind. They describe it by name and trust the Assistant to find it. Most of the time, it does. Sometimes, it doesn’t. And when it fails, it fails silently.</p><h3>Problems</h3><p>This surfaced two distinct problems :</p><ol><li><strong>Retrieval can silently fail on the exact document the user needs</strong></li></ol><p>If the true filename is template_contract_3.pdf but the user typed contract_template_3.pdf, retrieval may return nothing — or the wrong file. The same behavior is expected if the user targets a specific folder name.</p><p>Missing the exact document the user had in mind is a silent failure: the assistant gives no warning, and the user is left wondering why their question wasn’t answered properly.</p><p><strong>2. Users cannot know whether their references were picked up</strong></p><p>When a user references a document by its name or description, she has no confirmation that it was picked up correctly. Seeing the document explicitly flagged in the query, and being able to open it is reassuring, and it helps users be more precise.</p><p><em>To solve these problems, we introduced explicit document references, read on for the details…</em></p><p>Solution</p><p>We drew inspiration from existing interfaces such as Claude Code or Notion. Users can reference files, folders or pages directly in the input with “@”. This creates a familiar and intuitive way to attach context within the Assistant.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ry7-GefUc8IZVQ53UoweNw.png" /></figure><h3>1) Auto-complete for document and folder names</h3><p>Since we already index all the document names and folders of the users’ documents, it is quite straightforward to wire that index into a real-time autocomplete. As soon as the user types “@”, suggestions appear based on document names and folders.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*IlUk1_Bhgp8my9gsofGrBQ.png" /></figure><h3>2) A visually enriched query input</h3><p>Referenced documents are highlighted in blue within the query and shown as clickable chips above the input. At a glance, users can see which documents are referenced, and open them directly.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*c6b8W6HtCcYuWXlLsyTJpw.png" /></figure><h3>3) Guaranteed retrieval of reference documents</h3><p>Explicitly referenced documents are always included in the retrieval results. They’re added on top of the standard retrieved candidates, so the assistant never silently drops a document the user pointed to.</p><h3>Expected impact</h3><p>We expect users to stop describing documents in their queries and start referencing them instead. Seeing the document chip in the query will give users confidence that the Assistant understands them. And the silent retrieval failures on named documents should drop to zero, which contributes as much to the overall quality of our assistant than it does to the general user experience, a win-win for everyone.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c4fbf3a7ed95" width="1" height="1" alt=""><hr><p><a href="https://medium.com/doctrine/our-users-knew-exactly-which-documents-they-needed-our-assistant-didnt-here-s-what-we-built-c4fbf3a7ed95">Our users knew exactly which documents they needed. Our assistant didn’t. Here’s what we built.</a> was originally published in <a href="https://medium.com/doctrine">Inside Doctrine</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[ICDAR 2024: key insights from the future of Document Analysis]]></title>
            <link>https://medium.com/doctrine/icdar-2024-key-insights-from-the-future-of-document-analysis-9249d98a934a?source=rss-c887c489dd6a------2</link>
            <guid isPermaLink="false">https://medium.com/p/9249d98a934a</guid>
            <category><![CDATA[ocr]]></category>
            <category><![CDATA[engineering]]></category>
            <category><![CDATA[nlp]]></category>
            <category><![CDATA[conference]]></category>
            <category><![CDATA[document-analysis]]></category>
            <dc:creator><![CDATA[Maud Parrot]]></dc:creator>
            <pubDate>Wed, 19 Mar 2025 07:59:47 GMT</pubDate>
            <atom:updated>2025-03-19T07:59:47.403Z</atom:updated>
            <content:encoded><![CDATA[<p><em>by </em><a href="https://medium.com/@thibaud.rds"><em>Thibaud Real</em></a><em> and </em><a href="https://medium.com/@maud.parrot_925"><em>Maud Parrot</em></a></p><h3>Introduction</h3><p>Attending <strong>ICDAR 2024</strong> in Athens was an <strong>inspiring</strong> experience. Our team at Doctrine joined hundreds of researchers and industry experts to learn, connect and share our work.</p><p>The conference offered a rich <strong>variety of experiences</strong>, including: engaging discussions, diverse formats, and high-quality presentations.</p><p>In this article, we will share <strong>key insights</strong> and quickly dive into<strong> three specific papers</strong> that captured our attention.</p><p>We even presented <strong>our own research</strong> at one of the conference workshop! We’re breaking down that work in detail in a separate article [<a href="https://medium.com/doctrine/structuring-legal-documents-a-multimodal-framework-9b16b54e1813">Structuring Legal Documents: a Multimodal Framework</a>].</p><h3>Key Takeaways</h3><p>The conference format was <strong>quite diverse</strong>. From in-depth tutorials and technical workshops to poster sessions and oral presentations, each format served its purpose in <strong>sharing knowledge</strong>. The keynotes delivered valuable insights into the field’s direction, for example questioning the paradigm of a multi-step workflow (OCR, table detection, structure, …) vs a unified model performing all these steps at once.</p><p><strong>Layout analysis</strong> and <strong>document understanding</strong> dominated the conversations, with significant attention paid to <strong>handwritten recognition</strong> and <strong>table extraction</strong>. <strong>Historical document</strong> analysis and <strong>music recognition</strong> also emerged as interesting areas of research.</p><p>While <strong>transformer</strong> architectures are widely adopted, LLMs aren’t yet the go-to solution in this domain. Instead, researchers are focusing on <strong>hybrid approaches</strong> using Transformer models that consider <strong>both image and text modalities.</strong></p><p>The field’s approach to <strong>training data</strong> was also interesting. While <strong>synthetic datasets</strong> are increasingly common, researchers have also noticed <strong>their limitations</strong> in representing real-world complexity. This challenge was particularly present in tasks like OCR, where this year’s focus shifted more toward improving image quality through, for example, <strong>dewarping and noise reduction</strong> rather than core OCR advances.</p><p>Among all the rich presentations, we introduce below 3 papers from the conference that caught our attention.</p><h3>1. DLAFormer: An End-to-End Transformer for Document Layout Analysis</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*uec2zq3u3JwE7kqTOJummg.png" /></figure><h4><strong>Summary:</strong></h4><ul><li>Traditionally, document layout analysis has been a fragmented process, relying on separate models for tasks like table detection, figure extraction, and text region segmentation.</li><li>DLAFormer uses an <strong>end-to-end transformer-based</strong> architecture for document layout analysis to perform all these tasks as a <strong>relation prediction problem</strong></li><li>Blocs in the documents may have relation of multiple types : <strong>inter, intra, and logical role.</strong> Each part of the document can have multiple relation types.</li><li>The output is a <strong>single matrix</strong> that contains the <strong>unified label space</strong>, allowing the model to predict all relation types (inter, intra, and logical role) using a simplified output representation. That matrix can define preceding and following blocks, and also the block type (a table, a header, etc.).</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*0KgQrZsFYTTyW92oQzvBSQ.png" /></figure><h4><strong>Resources:</strong></h4><ul><li>Paper: <a href="https://arxiv.org/pdf/2405.11757">https://arxiv.org/pdf/2405.11757</a></li></ul><h3>2. GDP: Generic Document Pretraining to Improve Document Understanding</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*S8jMuQyXrliUNjIf9WnSOg.png" /></figure><h4>Summary:</h4><ul><li>GDPerformer is a self-supervised model designed for document understanding.</li><li>It pre-trains architectures to predict both <strong>masked OCR tokens</strong> and <strong>masked bounding boxes</strong>, allowing the network to learn rich document representations such as structure and language.</li><li><strong>Challenges</strong>: The model deals with <strong>handwritten text</strong>, <strong>unstructured documents</strong>, and <strong>noisy data</strong>.</li><li><strong>Key innovation</strong>: Unlike earlier models (e.g., <strong>LayTr</strong>, <strong>DocFormer</strong>), GDPerformer masks both tokens and bounding boxes, enhancing its understanding of document layouts.</li><li><strong>A second version GDPerformerv2</strong> incorporates <strong>image features</strong> to further improve performance.</li><li><strong>Results</strong>: After pre-training, GDPerformer is fine-tuned on tasks like <strong>entity recognition</strong> and <strong>classification</strong>, achieving <strong>state-of-the-art (SOTA)</strong> results on various benchmarks, including the <strong>Tobacco dataset</strong>. While <strong>V1</strong> performs well in semantic recognition, <strong>V2</strong> significantly improves performance, reaching SOTA results.</li></ul><h4>Resources:</h4><ul><li>Paper: <a href="https://www.springerprofessional.de/gdp-generic-document-pretraining-to-improve-document-understandi/27665094"><strong>GDP: Generic Document Pretraining to Improve Document Understanding</strong></a></li></ul><h3>3. ClusterTabNet: Supervised clustering method for table detection and table structure recognition</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Tm-oCkhf-4BYbL5xhdekRQ.png" /></figure><h4><strong>Summary:</strong></h4><ul><li>Objective: The model <strong>clusters words in OCR output</strong> to detect and recognize tables.</li><li>Table structure modeling : Represents table structure as a graph embedded in an <strong>adjacency matrix</strong> between words. Words are linked when belonging to the same row, column, header, as well as to the same table.</li><li>A Transformer encoder is trained to predict an <strong>adjacency matrix</strong></li><li><strong>Four different clustering tasks</strong>: table detection, header clustering, row clustering, and column clustering.</li><li>For each clustering target (tables, rows, columns, headers), the model outputs an adjacency matrix with size equal to the number of words in the page.</li><li><strong>Use of Embeddings</strong>: Each patch embedding is created from the OCR output and goes through a CNN, followed by a transformer encoder with a clustering head.</li><li><strong>Role of Keys and Values</strong>: The extracted keys and values are used as attention to improve extraction accuracy.</li><li><strong>Importance of the Image</strong>: The paper suggests that the image’s role in the table extraction process is not very significant.</li><li><strong>Limitations</strong>: It doesn’t handle tables that span multiple pages, and the output matrix size can be large. This system requires 4 different model inference :one for each task.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*HKRVd3RzN9r8Xe06iNaNJg.png" /></figure><h4><strong>Resources :</strong></h4><ul><li>Paper: <a href="https://arxiv.org/pdf/2402.07502">https://arxiv.org/pdf/2402.07502</a></li><li><a href="https://github.com/SAP-samples/clustertabnet/tree/main">Code</a></li></ul><h3>Conclusion: a hybrid future of document analysis</h3><p>ICDAR 2024 provided <strong>interesting insights</strong> into the rapidly evolving field of <strong>document analysis</strong>. The papers we have highlighted — DLAFormer, GDPerformer, and ClusterTabNet — present interesting trends of this field.</p><p>Notably, it seems that the document analysis evolves towards <strong>hybrid and unified end-to-end approaches.</strong> Integrating transformers with multimodal processing, is particularly present. These methods, which leverage both visual and textual information, seem to be the path for building more robust and accurate document understanding systems.</p><p>As we move forward, we anticipate further advancements in handling complex document structures, noisy data, and the nuances of real-world documents. The research presented at ICDAR 2024 has laid a strong foundation for these future developments, and we’re excited to see what the next chapter in document analysis will bring. We at <strong>Doctrine</strong> are committed to staying at the forefront of these innovations and leveraging them to deliver better products for our clients.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9249d98a934a" width="1" height="1" alt=""><hr><p><a href="https://medium.com/doctrine/icdar-2024-key-insights-from-the-future-of-document-analysis-9249d98a934a">ICDAR 2024: key insights from the future of Document Analysis</a> was originally published in <a href="https://medium.com/doctrine">Inside Doctrine</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>