# document ai

Document AI is a document processing and understanding platform that transforms unstructured document data into structured data and supports OCR, extraction, classification, and document-processing workflows.

This is one page of public article previews, not the complete archive. Follow Next page to continue. Summaries are not the original full articles.

## Data Extraction for AI Agents: 10 MCP Prompts for PDFs, Scans and Forms

DevFeed: [Data Extraction for AI Agents: 10 MCP Prompts for PDFs, Scans and Forms](<https://devfeed.tech/articles/data-extraction-for-ai-agents-10-mcp-prompts-for-pdfs-scans-and-forms-143904.md>)

Original publisher: [Read original article](<https://hackernoon.com/data-extraction-for-ai-agents-10-mcp-prompts-for-pdfs-scans-and-forms?source=rss>)

Author: Nutrient

Published: 2026-11-19T16:06:18Z

Content type: tutorial

Language: en

Sources: [HackerNoon](<https://devfeed.tech/sources/hackernoon.md>)

Topics: [document ai](<https://devfeed.tech/topics/document-ai.md>), [AI for data automation](<https://devfeed.tech/topics/ai-for-data-automation.md>), [Amazon Textract](<https://devfeed.tech/topics/amazon-textract.md>)

Tags: [agents](<https://devfeed.tech/tags/agents.md>), [ai](<https://devfeed.tech/tags/ai.md>), [ai-agent-mcp](<https://devfeed.tech/tags/ai-agent-mcp.md>), [ai-agents](<https://devfeed.tech/tags/ai-agents.md>), [ai-document-processing](<https://devfeed.tech/tags/ai-document-processing.md>), [api](<https://devfeed.tech/tags/api.md>), [batch](<https://devfeed.tech/tags/batch.md>), [claude](<https://devfeed.tech/tags/claude.md>), [claude-code](<https://devfeed.tech/tags/claude-code.md>), [codex](<https://devfeed.tech/tags/codex.md>), [codex-mcp](<https://devfeed.tech/tags/codex-mcp.md>), [confidence](<https://devfeed.tech/tags/confidence.md>), [contracts](<https://devfeed.tech/tags/contracts.md>), [cost](<https://devfeed.tech/tags/cost.md>), [csv](<https://devfeed.tech/tags/csv.md>), [data](<https://devfeed.tech/tags/data.md>), [data-extraction](<https://devfeed.tech/tags/data-extraction.md>), [document-processing](<https://devfeed.tech/tags/document-processing.md>), [forms](<https://devfeed.tech/tags/forms.md>), [llm-training](<https://devfeed.tech/tags/llm-training.md>), [llms](<https://devfeed.tech/tags/llms.md>), [mcp](<https://devfeed.tech/tags/mcp.md>), [pdf](<https://devfeed.tech/tags/pdf.md>), [pdfs](<https://devfeed.tech/tags/pdfs.md>), [prompts](<https://devfeed.tech/tags/prompts.md>), [tables](<https://devfeed.tech/tags/tables.md>)

### AI overview

The prompts connect Claude Code, Codex, or the Claude app to Nutrient's hosted MCP server so a document extraction API can return structured values with page citations, bounding boxes, and groundedness scores. In the tested invoice, all fields were correct, while a low-confidence credit-line price was flagged for human review. The examples also cover OCR for scans, table and batch extraction, Markdown parsing, classification, redaction, and local page-count cost estimates. Results have limits: citations can cover a source region rather than an exact word; multi-column PDFs may have reading-order and table errors; batch currency and date formats can vary or produce blanks; and redaction may miss signatures or remove extra data. The article gives rates of $0.015/page for extraction, $0.018 for agentic parsing, and $0.009 for understand-mode parsing, and recommends checking cited regions and reviewing consequential outputs.

### Source excerpt

10 copy-paste MCP prompts for AI agents: extract invoices, tables and scans with a data extraction API, with a page and bounding box for every value.

## What Is BIO Tagging and How to Use It for Document AI Training Data

DevFeed: [What Is BIO Tagging and How to Use It for Document AI Training Data](<https://devfeed.tech/articles/what-is-bio-tagging-and-how-to-use-it-for-document-ai-training-data-146784.md>)

Original publisher: [Read original article](<https://www.sitepoint.com/what-is-bio-tagging-and-how-to-use-it-for-document-ai-training-data/>)

Author: Bilal Ahmad

Published: 2026-11-19T12:24:33Z

Content type: tutorial

Language: en

Sources: [SitePoint](<https://devfeed.tech/sources/sitepoint.md>)

Topics: [document ai](<https://devfeed.tech/topics/document-ai.md>), [AI-Ready Data](<https://devfeed.tech/topics/ai-ready-data.md>), [Data Quality](<https://devfeed.tech/topics/data-quality.md>), [Machine learning](<https://devfeed.tech/topics/machine-learning.md>), [Advanced AI workflows](<https://devfeed.tech/topics/advanced-ai-workflows.md>), [Transformers](<https://devfeed.tech/topics/transformers.md>)

Tags: [ai](<https://devfeed.tech/tags/ai.md>), [ai-training](<https://devfeed.tech/tags/ai-training.md>), [ai-training-data](<https://devfeed.tech/tags/ai-training-data.md>), [bio](<https://devfeed.tech/tags/bio.md>), [data](<https://devfeed.tech/tags/data.md>), [document-ai](<https://devfeed.tech/tags/document-ai.md>)

### AI overview

BIO tagging labels each token as the beginning (B), continuation (I), or outside (O) of an entity. For document AI, the article pairs token labels with normalized bounding boxes so models such as LayoutLMv3 can learn both text and page position. It recommends defining a consistent, fine-grained label taxonomy before annotation, extracting words and coordinates from PDFs or OCR, assigning BIO labels, and formatting parallel words, boxes, and labels for training. Regex pre-labeling can cover some predictable fields, while active learning and auto-labeling platforms shift effort toward review; names, addresses, and line descriptions still need human judgment. Validate entity boundaries, BIO sequences, label balance, and bounding boxes before training. The article suggests starting with 50-100 labeled documents, fine-tuning a pretrained checkpoint, then checking errors before scaling; its stated accuracy ranges depend on document diversity and assume pretrained fine-tuning.

### Source excerpt

null Continue reading What Is BIO Tagging and How to Use It for Document AI Training Data on SitePoint.

## Nutrient Data Extraction API: JSON Matching Your Schema With Per-Field Citations and Confidence

DevFeed: [Nutrient Data Extraction API: JSON Matching Your Schema With Per-Field Citations and Confidence](<https://devfeed.tech/articles/nutrient-data-extraction-api-json-matching-your-schema-with-per-field-citations-and-confidence-142244.md>)

Original publisher: [Read original article](<https://www.marktechpost.com/2026/11/18/nutrient-data-extraction-api-citations-confidence/>)

Author: Asif Razzaq

Published: 2026-11-19T03:23:50Z

Content type: article

Language: en

Sources: [MarkTechPost](<https://devfeed.tech/sources/marktechpost.md>)

Topics: [document ai](<https://devfeed.tech/topics/document-ai.md>), [API](<https://devfeed.tech/topics/api.md>), [JSON](<https://devfeed.tech/topics/json.md>), [schema design](<https://devfeed.tech/topics/schema-design.md>)

Tags: [agentic-ai](<https://devfeed.tech/tags/agentic-ai.md>), [ai-shorts](<https://devfeed.tech/tags/ai-shorts.md>), [api](<https://devfeed.tech/tags/api.md>), [applications](<https://devfeed.tech/tags/applications.md>), [artificial-intelligence](<https://devfeed.tech/tags/artificial-intelligence.md>), [computer-vision](<https://devfeed.tech/tags/computer-vision.md>), [confidence](<https://devfeed.tech/tags/confidence.md>), [data-extraction](<https://devfeed.tech/tags/data-extraction.md>), [editors-pick](<https://devfeed.tech/tags/editors-pick.md>), [embedding-model](<https://devfeed.tech/tags/embedding-model.md>), [enterprise-ai](<https://devfeed.tech/tags/enterprise-ai.md>), [for-devs](<https://devfeed.tech/tags/for-devs.md>), [generative-ai](<https://devfeed.tech/tags/generative-ai.md>), [json](<https://devfeed.tech/tags/json.md>), [language-model](<https://devfeed.tech/tags/language-model.md>), [new-releases](<https://devfeed.tech/tags/new-releases.md>), [ocr](<https://devfeed.tech/tags/ocr.md>), [promote](<https://devfeed.tech/tags/promote.md>), [schema](<https://devfeed.tech/tags/schema.md>), [sponsored](<https://devfeed.tech/tags/sponsored.md>), [staff](<https://devfeed.tech/tags/staff.md>), [studio](<https://devfeed.tech/tags/studio.md>), [technology](<https://devfeed.tech/tags/technology.md>), [test](<https://devfeed.tech/tags/test.md>), [uncategorized](<https://devfeed.tech/tags/uncategorized.md>), [vision-language-model](<https://devfeed.tech/tags/vision-language-model.md>)

### AI overview

Nutrient's Extraction API maps PDFs, scans, images, and Office files to a supplied JSON schema, attaching page and region citations, match labels, and confidence when available. Studio lets teams draft schemas, inspect results on fresh documents, save versioned processors, and evaluate them against up to 25 reviewed documents. Parsing modes include structure, understand, and agentic; Auto routes pages among them, while omitted mode defaults to understand. Confidence combines optional signals and can be missing; even a grounded value may answer the wrong field, so the article's example checks match labels before confidence and still calls for type and business-rule validation. The service is available in Nutrient's cloud, through an SDK on customer servers, or on Apple silicon; fully offline production requires the SDK. Nutrient reports an 86.2 WholeDoc Bench score for agentic mode, but says the benchmark measures parsing rather than extraction accuracy, its main test set is private, and cross-page tables and charts remain difficult.

### Source excerpt

Nutrient Data Extraction API parses PDFs, scans and Office files into JSON matching your schema. Fields carry page, region and match-label citations, and Studio lets teams test before writing code. The post Nutrient Data Extraction API: JSON Matching Your Schema With Per-Field Citations and Confidence appeared first on MarkTechPost.

## Unsiloed nabs $7.5M to turn scanned documents into structured data for AI agents

DevFeed: [Unsiloed nabs $7.5M to turn scanned documents into structured data for AI agents](<https://devfeed.tech/articles/unsiloed-nabs-7-5m-to-turn-scanned-documents-into-structured-data-for-ai-agents-140862.md>)

Original publisher: [Read original article](<https://siliconangle.com/2026/11/18/unsiloed-nabs-7-5m-to-turn-scanned-documents-into-structured-data-for-ai-agents/>)

Author: Kyt Dotson

Published: 2026-11-18T16:00:40Z

Content type: news

Language: en

Sources: [SiliconANGLE](<https://devfeed.tech/sources/siliconangle.md>)

Topics: [document ai](<https://devfeed.tech/topics/document-ai.md>), [Amazon Textract](<https://devfeed.tech/topics/amazon-textract.md>), [AI-Ready Data](<https://devfeed.tech/topics/ai-ready-data.md>), [Serialization protocols](<https://devfeed.tech/topics/serialization-protocols.md>), [Artificial Intelligence](<https://devfeed.tech/topics/ai.md>)

Tags: [ai](<https://devfeed.tech/tags/ai.md>), [ai-agents](<https://devfeed.tech/tags/ai-agents.md>), [ai-startup-funding](<https://devfeed.tech/tags/ai-startup-funding.md>), [aman-mishra](<https://devfeed.tech/tags/aman-mishra.md>), [apps](<https://devfeed.tech/tags/apps.md>), [artificial-intelligence](<https://devfeed.tech/tags/artificial-intelligence.md>), [computer-vision](<https://devfeed.tech/tags/computer-vision.md>), [funding](<https://devfeed.tech/tags/funding.md>), [news](<https://devfeed.tech/tags/news.md>), [pdfs](<https://devfeed.tech/tags/pdfs.md>), [scanning](<https://devfeed.tech/tags/scanning.md>), [startup-funding](<https://devfeed.tech/tags/startup-funding.md>), [structured-data](<https://devfeed.tech/tags/structured-data.md>), [the-latest](<https://devfeed.tech/tags/the-latest.md>), [unsiloed](<https://devfeed.tech/tags/unsiloed.md>)

### AI overview

Unsiloed says its visual-first parser converts PDFs, forms, presentations, Word files and images into Markdown or JSON while preserving layout, tables, sections and relationships between content. Unlike OCR, which the article describes as extracting text without recognizing layout or graphics, Unsiloed combines computer vision algorithms with dual-stream vision models that process text and visual structure together. Developers can provide schemas to extract fields, classify or split documents, and update text and images while retaining formatting. The company reports an 88.0 score on olmOCR-Bench, compared with 81.9 for Claude Opus 4.7, 84.6 for GPT 5.5 and 77.7 for Gemini 3 Pro, and says it ranked first among 19 systems. It supports private cloud, on-premises and air-gapped deployment. The excerpt reports a $7.5 million seed round led by Felicis.

### Source excerpt

Document-scanning firm Unsiloed Inc. today announced it has raised $7.5 million in seed funding to turn PDFs and other complex documents into structured, readable data for artificial intelligence agents. Participants in the round, led by Felicis, included Y Combinator, Orange Collective, Uncorrelated Ventures, Aarthi Ramamurthy and Beth Turner. A variety of documents arrive in formats [...] The post Unsiloed nabs $7.5M to turn scanned documents into structured data for AI agents appeared first on SiliconANGLE.

## Felicis leads $7.5M bet on Unsiloed AI to make enterprise documents usable by AI agents

DevFeed: [Felicis leads $7.5M bet on Unsiloed AI to make enterprise documents usable by AI agents](<https://devfeed.tech/articles/felicis-leads-7-5m-bet-on-unsiloed-ai-to-make-enterprise-documents-usable-by-ai-agents-140573.md>)

Original publisher: [Read original article](<https://techcrunch.com/2026/11/18/felicis-leads-7-5m-bet-on-unsiloed-ai-to-make-enterprise-documents-usable-by-ai-agents/>)

Author: Jagmeet Singh

Published: 2026-11-18T12:00:00Z

Content type: news

Language: en

Sources: [TechCrunch | AI News & Artificial Intelligence](<https://devfeed.tech/sources/techcrunch-ai-news-artificial-intelligence.md>)

Topics: [document ai](<https://devfeed.tech/topics/document-ai.md>), [Artificial Intelligence](<https://devfeed.tech/topics/ai.md>), [data-processing](<https://devfeed.tech/topics/data-processing.md>), [AI Models](<https://devfeed.tech/topics/ai-models.md>), [agentic workflows](<https://devfeed.tech/topics/agentic-workflows.md>)

Tags: [agents](<https://devfeed.tech/tags/agents.md>), [ai](<https://devfeed.tech/tags/ai.md>), [ai-agents](<https://devfeed.tech/tags/ai-agents.md>), [ai-document-processing](<https://devfeed.tech/tags/ai-document-processing.md>), [california](<https://devfeed.tech/tags/california.md>), [document-processing](<https://devfeed.tech/tags/document-processing.md>), [enterprise](<https://devfeed.tech/tags/enterprise.md>), [exclusive](<https://devfeed.tech/tags/exclusive.md>), [felicis](<https://devfeed.tech/tags/felicis.md>), [felicis-ventures](<https://devfeed.tech/tags/felicis-ventures.md>), [startups](<https://devfeed.tech/tags/startups.md>), [unsiloed](<https://devfeed.tech/tags/unsiloed.md>), [y-combinator](<https://devfeed.tech/tags/y-combinator.md>)

### AI overview

Unsiloed AI converts complex business documents into structured data for AI agents and links extracted information back to its source so users can verify it. The company says its dual-stream vision models process text and images separately, use heat-map chunking to focus computing on key areas, and apply an industry-specific decoder. It reports 99.78% accuracy across about 84,000 data points from 542 financial documents in a customer test, and an 88.0 score on olmOCR-Bench versus 66.0 for Reducto and 39.9 for Unstructured. These results were produced or shared by the startup and have not been independently verified. Unsiloed says pricing starts at one cent per page; it has customers in finance, legal services, and healthcare.

### Source excerpt

With a team of just seven, Unsiloed AI plans to compete with tech giants and specialized startups in the crowded AI document-processing market.

## Building a Private AI Document Assistant on an 8GB GPU Using RAG and llama.cpp

DevFeed: [Building a Private AI Document Assistant on an 8GB GPU Using RAG and llama.cpp](<https://devfeed.tech/articles/building-a-private-ai-document-assistant-on-an-8gb-gpu-using-rag-and-llama-cpp-138032.md>)

Original publisher: [Read original article](<https://hackernoon.com/building-a-private-ai-document-assistant-on-an-8gb-gpu-using-rag-and-llamacpp?source=rss>)

Author: Naresh Waghela

Published: 2026-11-17T11:45:29Z

Content type: tutorial

Language: en

Sources: [HackerNoon](<https://devfeed.tech/sources/hackernoon.md>)

Topics: [llama.cpp](<https://devfeed.tech/topics/llama-cpp.md>), [Artificial Intelligence](<https://devfeed.tech/topics/ai.md>), [AI Chat](<https://devfeed.tech/topics/ai-chat.md>), [document ai](<https://devfeed.tech/topics/document-ai.md>)

Tags: [ai](<https://devfeed.tech/tags/ai.md>), [ai-engineering](<https://devfeed.tech/tags/ai-engineering.md>), [artificial-intelligence](<https://devfeed.tech/tags/artificial-intelligence.md>), [databases](<https://devfeed.tech/tags/databases.md>), [faiss](<https://devfeed.tech/tags/faiss.md>), [gpu](<https://devfeed.tech/tags/gpu.md>), [large-language-models](<https://devfeed.tech/tags/large-language-models.md>), [llama-cpp](<https://devfeed.tech/tags/llama-cpp.md>), [local-llm](<https://devfeed.tech/tags/local-llm.md>), [private](<https://devfeed.tech/tags/private.md>), [private-ai](<https://devfeed.tech/tags/private-ai.md>), [private-ai-assistant](<https://devfeed.tech/tags/private-ai-assistant.md>), [python](<https://devfeed.tech/tags/python.md>), [python-rag](<https://devfeed.tech/tags/python-rag.md>), [rag](<https://devfeed.tech/tags/rag.md>)

### AI overview

The article presents a local RAG document assistant using Qwen2.5-7B-Instruct Q4_K_M with llama.cpp on an 8GB GPU, while Sentence Transformers embeddings and FAISS retrieval run on the CPU. It extracts text from PDFs, Markdown, and TXT files, stores overlapping chunks with file and page IDs, retrieves passages, and checks the assembled prompt with llama.cpp's tokenizer before generation. The CLI shows answers, retrieved sources, and timings; a fixed-question benchmark supports repeatable evaluation, but the article reports no measured GPU results. It also explains limits: scanned PDFs need OCR, extraction and retrieval can fail, citations require verification, and local execution alone does not ensure security. The reference app is a single-user baseline without authentication or document permissions.

### Source excerpt

Build a private AI document assistant on an 8GB GPU using llama.cpp, Python, FAISS and RAG. Includes working code, citations and reproducible tests.

## Benchmarking Multilingual Document Understanding in Vision-Language Models

DevFeed: [Benchmarking Multilingual Document Understanding in Vision-Language Models](<https://devfeed.tech/articles/benchmarking-multilingual-document-understanding-in-vision-language-models-132044.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.14758>)

Author: Matteo Gabburo, Daniela Baiamonte, Stefano Simonazzi, Silvia Gardini, Sean Cesare Lazzeri, Andrea Zugarini

Published: 2026-11-13T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>)

Topics: [LLM evaluation / benchmarking](<https://devfeed.tech/topics/llm-evaluation-benchmarking.md>), [vlm](<https://devfeed.tech/topics/vlm.md>), [document ai](<https://devfeed.tech/topics/document-ai.md>)

Tags: [benchmark](<https://devfeed.tech/tags/benchmark.md>), [benchmarking](<https://devfeed.tech/tags/benchmarking.md>), [benchmarks](<https://devfeed.tech/tags/benchmarks.md>), [capabilities](<https://devfeed.tech/tags/capabilities.md>), [classification](<https://devfeed.tech/tags/classification.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-cl](<https://devfeed.tech/tags/cs-cl.md>), [domain](<https://devfeed.tech/tags/domain.md>), [english](<https://devfeed.tech/tags/english.md>), [examples](<https://devfeed.tech/tags/examples.md>), [gap](<https://devfeed.tech/tags/gap.md>), [language](<https://devfeed.tech/tags/language.md>), [language-models](<https://devfeed.tech/tags/language-models.md>), [languages](<https://devfeed.tech/tags/languages.md>), [layout](<https://devfeed.tech/tags/layout.md>), [layouts](<https://devfeed.tech/tags/layouts.md>), [markdown](<https://devfeed.tech/tags/markdown.md>), [model](<https://devfeed.tech/tags/model.md>), [models](<https://devfeed.tech/tags/models.md>), [multimodal](<https://devfeed.tech/tags/multimodal.md>), [vision](<https://devfeed.tech/tags/vision.md>), [vision-language-models](<https://devfeed.tech/tags/vision-language-models.md>)

### AI overview

The MDU benchmark finds that multilingual document understanding remains uneven across languages and tasks, and larger models do not reliably solve those gaps. It contains 23,146 examples across 13 tasks in English, French, German, Italian, and Spanish. In evaluations of 45 open-weight VLMs (1.7B-235B parameters) and one commercial system under the same protocol, the top average was 69.0; only eight open-weight models scored at least 65, and the two highest results came from 27B models. The strongest-to-weakest language gap persisted as model size increased and after controlling for language domain composition. Task abilities were largely uncorrelated after accounting for overall model quality, with transcription in particular trading off against semantic extraction.

### Source excerpt

arXiv:2611.14758v1 Announce Type: cross Abstract: Recent Vision-Language Models (VLMs) perform strongly on multimodal tasks, Document Understanding (DU) among them, but the benchmarks that measure them mostly cover English documents and narrow problems. How well current VLMs generalize across languages, tasks, styles and layouts therefore remains unclear. We introduce MDU, a multilingual DU benchmark of 23146 examples spanning 13 tasks and five languages: English, French, German, Italian, and Spanish. The tasks isolate complementary capabilities, from surface-level text and Markdown extraction through summarization, classification and named-entity recognition to layout-sensitive passage and layout extraction. We evaluate 45 state-of-the-art open-weight VLMs from 1.7B to 235B parameters, plus one commercial system under the same protocol. The benchmark is far from saturated: the best average is 69.0, only eight of the 45 evaluated systems reach 65, and the two best results come from 27B models rather than from the largest systems we evaluate. Performance is uneven in ways that scale does not fix. The gap between the strongest and weakest language does not close as models grow, and it survives controlling for the domain composition of each language. Capabilities also do not improve together: once we partial out overall model quality, per-task scores are largely uncorrelated, and transcription ability in particular trades off against semantic extraction.

## Read, Parse, Describe: Unified Document Parsing with Visual Element Description Generation

DevFeed: [Read, Parse, Describe: Unified Document Parsing with Visual Element Description Generation](<https://devfeed.tech/articles/read-parse-describe-unified-document-parsing-with-visual-element-description-generation-131992.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2610.20335>)

Author: Yuheng Chen, Yufan Chen, Zhuojun Cai, Ruiping Liu, Junwei Zheng, Jiale Wei, Di Wen, Weijia Fan, Kunyu Peng, Jiaming Zhang, Rainer Stiefelhagen

Published: 2026-11-13T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>)

Topics: [document ai](<https://devfeed.tech/topics/document-ai.md>), [alt text](<https://devfeed.tech/topics/alt-text.md>)

Tags: [benchmark](<https://devfeed.tech/tags/benchmark.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-cv](<https://devfeed.tech/tags/cs-cv.md>), [dataset](<https://devfeed.tech/tags/dataset.md>), [grid](<https://devfeed.tech/tags/grid.md>), [guidance](<https://devfeed.tech/tags/guidance.md>), [language](<https://devfeed.tech/tags/language.md>), [layout](<https://devfeed.tech/tags/layout.md>), [order](<https://devfeed.tech/tags/order.md>), [parsing](<https://devfeed.tech/tags/parsing.md>), [reading](<https://devfeed.tech/tags/reading.md>), [spatial](<https://devfeed.tech/tags/spatial.md>), [vision](<https://devfeed.tech/tags/vision.md>), [vlms](<https://devfeed.tech/tags/vlms.md>)

### AI overview

Read-Parse-Describe benchmarks unified document parsing from page images: models produce ordered layout elements with bounding boxes, semantic labels, reading order, recognized text, and alternative descriptions of visual elements. Its dataset covers 2,402 documents and 18,163 pages, with 125,218 layout elements and 11,338 visual instances. The authors also propose a plug-and-play space-awareness enhancer that reuses multi-layer visual features and adds grid-aware spatial guidance to vision-language tokens. Experiments with end-to-end VLMs and pipeline parsers find that current systems struggle to jointly recover reading order and visual semantics; the enhanced model is reported to achieve strong joint performance.

### Source excerpt

arXiv:2610.20335v2 Announce Type: replace-cross Abstract: Document understanding systems are typically evaluated as separate components via layout analysis, optical text recognition, reading-order prediction, and image captioning. This separation leaves scientific figures, captions, and visual semantics weakly connected, and fails to measure whether a model can produce a single page-level parse that is both structurally faithful and semantically accessible. We introduce Read-Parse-Describe, a unified dataset for document parsing with visual element description generation. Given a page image, models are required to output an ordered list of layout elements containing bounding boxes, semantic labels, reading-order indices, recognized textual content, and alternative text descriptions for visual elements. The benchmark is built from 2,402 documents, comprising 18,163 continuous pages, 125,218 layout elements, and 11,338 visual instances. Besides, we further propose a plug-and-play space-awareness enhancer that reuses multi-layer visual features and injects grid-aware spatial guidance into vision-language tokens for layout-aware generation. Experiments across end-to-end VLMs and pipeline parsers show that existing systems still struggle to jointly recover reading order and visual semantics, while our enhanced model achieves strong joint performance. These results highlight the need to move beyond isolated document parsing modules toward unified models that jointly read, parse, and describe complex documents. The new dataset and code will be made publicly available at https://github.com/yuheng98chen/RPD-Unidoc

## 138 Open Source OCR and PDF Parsers Ranked by GitHub Stars: PaddleOCR, MinerU, Tesseract and More

DevFeed: [138 Open Source OCR and PDF Parsers Ranked by GitHub Stars: PaddleOCR, MinerU, Tesseract and More](<https://devfeed.tech/articles/138-open-source-ocr-and-pdf-parsers-ranked-by-github-stars-paddleocr-mineru-tesseract-and-more-131028.md>)

Original publisher: [Read original article](<https://hackernoon.com/138-open-source-ocr-and-pdf-parsers-ranked-by-github-stars-paddleocr-mineru-tesseract-and-more?source=rss>)

Author: Text Mining

Published: 2026-11-12T17:52:14Z

Content type: article

Language: en

Sources: [HackerNoon](<https://devfeed.tech/sources/hackernoon.md>)

Topics: [document ai](<https://devfeed.tech/topics/document-ai.md>), [Open Source](<https://devfeed.tech/topics/open-source.md>), [pdf](<https://devfeed.tech/topics/pdf.md>)

Tags: [2026](<https://devfeed.tech/tags/2026.md>), [ai](<https://devfeed.tech/tags/ai.md>), [ai-ready-data](<https://devfeed.tech/tags/ai-ready-data.md>), [apache-2-0](<https://devfeed.tech/tags/apache-2-0.md>), [api](<https://devfeed.tech/tags/api.md>), [apple](<https://devfeed.tech/tags/apple.md>), [directory](<https://devfeed.tech/tags/directory.md>), [github](<https://devfeed.tech/tags/github.md>), [github-stars](<https://devfeed.tech/tags/github-stars.md>), [hackernoon-top-story](<https://devfeed.tech/tags/hackernoon-top-story.md>), [mining](<https://devfeed.tech/tags/mining.md>), [ocr](<https://devfeed.tech/tags/ocr.md>), [open](<https://devfeed.tech/tags/open.md>), [open-source](<https://devfeed.tech/tags/open-source.md>), [open-source-ocr](<https://devfeed.tech/tags/open-source-ocr.md>), [parsing](<https://devfeed.tech/tags/parsing.md>), [pdf](<https://devfeed.tech/tags/pdf.md>), [pdf-parsers](<https://devfeed.tech/tags/pdf-parsers.md>), [tesseract](<https://devfeed.tech/tags/tesseract.md>), [text](<https://devfeed.tech/tags/text.md>), [text-mining](<https://devfeed.tech/tags/text-mining.md>), [text-mining-github-stars](<https://devfeed.tech/tags/text-mining-github-stars.md>), [text-mining-ranked](<https://devfeed.tech/tags/text-mining-ranked.md>), [text-mining-ranked-by-github](<https://devfeed.tech/tags/text-mining-ranked-by-github.md>), [text-mining-ranked-by-stars](<https://devfeed.tech/tags/text-mining-ranked-by-stars.md>)

### AI overview

The directory ranks 138 open-source OCR and document-parsing repositories with at least 1,000 GitHub stars, based on a November 11, 2026 snapshot. PaddleOCR leads with 92.6K stars, followed by MinerU (83.5K) and Tesseract (77.5K). The list was selected from 321 tagged projects; 23 were excluded as not being OCR or document tools. It groups options by task, including PDF-to-Markdown or JSON for RAG, classic OCR, browser OCR, searchable scanned PDFs, tables and handwriting, math-to-LaTeX, screen capture, and document management. The text describes a shift toward converting PDFs into LLM-ready structured data and notes a tradeoff: template-based OCR can fail when layouts change, while vision LLMs handle varied documents more flexibly but cost more at volume; it suggests a hybrid approach. The broader category is mostly Python, and its licenses include permissive and copyleft options, so the source advises checking licenses before shipping.

### Source excerpt

PaddleOCR, MinerU and Tesseract lead 138 open source OCR and PDF-parsing repos with 1,000+ GitHub stars, ranked from the Text Mining directory.

## Build a Knowledge Layer: Turn Your Existing Data into Context

DevFeed: [Build a Knowledge Layer: Turn Your Existing Data into Context](<https://devfeed.tech/articles/build-a-knowledge-layer-turn-your-existing-data-into-context-127288.md>)

Original publisher: [Read original article](<https://neo4j.com/blog/graph-database/build-a-knowledge-layer-turn-your-existing-data-into-context/>)

Author: corydon baylor

Published: 2026-11-10T14:00:00Z

Content type: article

Language: en

Sources: [Neo4j](<https://devfeed.tech/sources/graph-database-technology-neo4j-blog.md>)

Topics: [knowledge-engineering](<https://devfeed.tech/topics/knowledge-engineering.md>), [AI-Ready Data](<https://devfeed.tech/topics/ai-ready-data.md>), [document ai](<https://devfeed.tech/topics/document-ai.md>), [AI Integration Strategies](<https://devfeed.tech/topics/ai-integration-strategies.md>), [semantic-layer](<https://devfeed.tech/topics/semantic-layer.md>), [Data pipelines](<https://devfeed.tech/topics/data-pipelines.md>)

Tags: [agents](<https://devfeed.tech/tags/agents.md>), [ai](<https://devfeed.tech/tags/ai.md>), [ai-agents](<https://devfeed.tech/tags/ai-agents.md>), [build](<https://devfeed.tech/tags/build.md>), [business](<https://devfeed.tech/tags/business.md>), [context](<https://devfeed.tech/tags/context.md>), [cypher](<https://devfeed.tech/tags/cypher.md>), [data](<https://devfeed.tech/tags/data.md>), [database](<https://devfeed.tech/tags/database.md>), [databricks](<https://devfeed.tech/tags/databricks.md>), [decision](<https://devfeed.tech/tags/decision.md>), [document-intelligence](<https://devfeed.tech/tags/document-intelligence.md>), [graph](<https://devfeed.tech/tags/graph.md>), [graph-database](<https://devfeed.tech/tags/graph-database.md>), [knowledge](<https://devfeed.tech/tags/knowledge.md>), [knowledge-layer](<https://devfeed.tech/tags/knowledge-layer.md>), [sources](<https://devfeed.tech/tags/sources.md>)

### AI overview

Document Intelligence turns unstructured files such as PDFs, Word files, and plain text into a knowledge graph through a guided prompt and optional refinement, without manual extraction or code. In the example, it extracts business owners, ownership shares, and a $400,000 loan request from Rosa's plan. Virtual Graph models existing warehouse tables as a graph without copying or migrating data; it translates Cypher queries into SQL run in the warehouse, where the data stays. Its model can be defined manually, generated from the database schema, or created with AI from table metadata, then refined in a point-and-click editor. The excerpt presents the two graphs as connected context for an agent assessing the loan; linking them is described as the subject of the next post.

### Source excerpt

In this series 1. Give Your AI Agents the Shared Context to Succeed 2. Map Your Data with an Ontology 3. Turn Your Existing Data into Context 4. Connect Your Data Sources with a Composite Database 5. Ground Decisions in... Read more ->

## RankShift: Evaluating Ranking Instability Across Benchmark Dimensions in Document Parsing

DevFeed: [RankShift: Evaluating Ranking Instability Across Benchmark Dimensions in Document Parsing](<https://devfeed.tech/articles/rankshift-evaluating-ranking-instability-across-benchmark-dimensions-in-document-parsing-125991.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.10679>)

Author: Halimat Afolabi, Daeyoung Lee

Published: 2026-11-10T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>)

Topics: [Ground truth / benchmark quality](<https://devfeed.tech/topics/ground-truth-benchmark-quality.md>), [document ai](<https://devfeed.tech/topics/document-ai.md>), [LLM evaluation / benchmarking](<https://devfeed.tech/topics/llm-evaluation-benchmarking.md>)

Tags: [benchmark](<https://devfeed.tech/tags/benchmark.md>), [bleu](<https://devfeed.tech/tags/bleu.md>), [composition](<https://devfeed.tech/tags/composition.md>), [correlation](<https://devfeed.tech/tags/correlation.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [dataset](<https://devfeed.tech/tags/dataset.md>), [design](<https://devfeed.tech/tags/design.md>), [format](<https://devfeed.tech/tags/format.md>), [interaction](<https://devfeed.tech/tags/interaction.md>), [json](<https://devfeed.tech/tags/json.md>), [leaderboard](<https://devfeed.tech/tags/leaderboard.md>), [markdown](<https://devfeed.tech/tags/markdown.md>), [matching](<https://devfeed.tech/tags/matching.md>), [meteor](<https://devfeed.tech/tags/meteor.md>), [metric](<https://devfeed.tech/tags/metric.md>), [model](<https://devfeed.tech/tags/model.md>), [models](<https://devfeed.tech/tags/models.md>), [ocr](<https://devfeed.tech/tags/ocr.md>), [pages](<https://devfeed.tech/tags/pages.md>), [parsing](<https://devfeed.tech/tags/parsing.md>), [ranking](<https://devfeed.tech/tags/ranking.md>)

### AI overview

Rankings of document parsing models can change substantially with benchmark design, especially when matching strategy interacts with ground-truth format. In RankShift's study of 16 OCR and vision-language models on OmniDocBench, JSON versus Markdown annotations produced mean Kendall's tau of about 0.33 with region-level matching, compared with 0.77 when each page was scored as one concatenated string. Dataset composition also shifted rankings (tau 0.48 across page-type strata and 0.49 across layout-type strata), while comparing OmniDocBench with Real5's reconstruction of the same pages was more stable (tau 0.80). Metric choice was comparatively stable: mean tau was about 0.78 across NED, BLEU, and METEOR. The authors argue leaderboards should report ranking robustness alongside headline scores.

### Source excerpt

arXiv:2611.10679v1 Announce Type: new Abstract: Document parsing benchmarks are often treated as stable measures of model quality, yet rankings may depend substantially on benchmark design choices. In this work, we present RankShift, a systematic study comparing OCR and vision-language models (VLMs) under controlled variation in dataset composition, matching strategy, and scoring methodology. Using OmniDocBench as a case study, we evaluate the ranking stability of 16 models across 3 dataset configurations, 6 matching strategies, and 3 scoring metrics. We quantify rank stability using pairwise rank correlation, top-k churn, and pairwise reversal rate. Our results show that the interaction between matching strategy and ground truth format is a substantial but understudied driver of instability: scoring against OmniDocBench's JSON annotations versus its Markdown rendering yields a mean Kendall's tau of approximately 0.33 under region-level matching strategies, with individual pairs as low as tau = 0.23, compared with tau = 0.77 when each page is scored as a single concatenated string. Dataset composition produces comparable instability across page-type strata (mean tau = 0.48) and layout-type strata (mean tau = 0.49), while the cross-dataset shift from OmniDocBench to Real5's Scanning reconstruction of the same pages is more stable (tau = 0.80). Metric choice is the most stable dimension, with mean tau of approximately 0.78 across NED, BLEU, and METEOR. These findings indicate that leaderboard position is not solely a property of model capability, motivating leaderboards that report not only headline scores but also the robustness of rankings across reasonable design choices.

## Entity-Aware Evaluation of OCR Systems for Historical Document Transcription

DevFeed: [Entity-Aware Evaluation of OCR Systems for Historical Document Transcription](<https://devfeed.tech/articles/entity-aware-evaluation-of-ocr-systems-for-historical-document-transcription-118285.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2610.33004>)

Author: Camilo Mari\\~no, Diego Belzarena, Marina Gardella, Ignacio Ram\\'irez, Gregory Randall, Jean-Michel Morel

Published: 2026-11-05T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>)

Topics: [document ai](<https://devfeed.tech/topics/document-ai.md>), [Data Quality](<https://devfeed.tech/topics/data-quality.md>), [ner](<https://devfeed.tech/topics/ner.md>), [Machine learning](<https://devfeed.tech/topics/machine-learning.md>)

Tags: [analysis](<https://devfeed.tech/tags/analysis.md>), [archive](<https://devfeed.tech/tags/archive.md>), [benchmark](<https://devfeed.tech/tags/benchmark.md>), [coverage](<https://devfeed.tech/tags/coverage.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-cl](<https://devfeed.tech/tags/cs-cl.md>), [cs-cv](<https://devfeed.tech/tags/cs-cv.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [eess-iv](<https://devfeed.tech/tags/eess-iv.md>), [error](<https://devfeed.tech/tags/error.md>), [errors](<https://devfeed.tech/tags/errors.md>), [evaluation](<https://devfeed.tech/tags/evaluation.md>), [language-models](<https://devfeed.tech/tags/language-models.md>), [matching](<https://devfeed.tech/tags/matching.md>), [metrics](<https://devfeed.tech/tags/metrics.md>), [ocr](<https://devfeed.tech/tags/ocr.md>), [optical-character-recognition](<https://devfeed.tech/tags/optical-character-recognition.md>), [quality](<https://devfeed.tech/tags/quality.md>), [qwen](<https://devfeed.tech/tags/qwen.md>), [rankings](<https://devfeed.tech/tags/rankings.md>), [systems](<https://devfeed.tech/tags/systems.md>), [transcription](<https://devfeed.tech/tags/transcription.md>), [vision-language-models](<https://devfeed.tech/tags/vision-language-models.md>)

### AI overview

The study finds that character error rate alone can miss whether OCR preserves historically meaningful entities such as names, dates, and identifiers. On 175 Berrutti Archive documents with 5,209 manually annotated entity mentions, it compares sequence-recognition OCR systems with OCR-specialized and general-purpose vision-language models, measuring entity coverage and recoverability at different character-level edit tolerances. Qwen 3.6 has the highest entity coverage at every tested tolerance. Several systems with lower exact-match coverage recover substantially when small errors are allowed, while others remain unable to recover many entities even with relaxed matching. The results show that OCR rankings depend on the evaluation criterion and that aggregate transcription accuracy does not capture entity preservation. The annotations are stated to be released upon acceptance.

### Source excerpt

arXiv:2610.33004v3 Announce Type: replace-cross Abstract: Optical Character Recognition (OCR) is essential for digitizing historical archives, yet standard evaluation metrics such as Character Error Rate (CER) do not necessarily reflect whether meaningful information, such as names, dates, and identifiers, is preserved. In this work, we investigate entitylevel analysis of OCR errors as a complementary perspective for evaluating OCR quality in historical documents. Using a publicly available benchmark of 175 documents from the Berrutti Archive, augmented with 5,209 manually annotated entity mentions that will be released upon acceptance, we compare sequence-recognition OCR systems, OCR-specialized vision-language models, and general-purpose vision-language models. Beyond global transcription accuracy, we analyze entity preservation and recoverability under varying character-level edit tolerances. Our experiments show that OCR rankings depend on the evaluation criterion: while Qwen 3.6 achieves the highest entity coverage across all tested tolerances, several systems with lower exact-match coverage recover substantially when small errors are tolerated, suggesting that many of their errors correspond to localized character-level deviations rather than fundamentally incorrect entity transcriptions. Conversely, some systems remain unable to recover many entities even under relaxed matching. These findings show that OCR quality cannot be fully characterized by aggregate transcription metrics alone and highlight the importance of evaluating how OCR errors affect the preservation and recoverability of meaningful information.

## Dharma Field Extractor: Any Document In, Structured Fields in JSON Out

DevFeed: [Dharma Field Extractor: Any Document In, Structured Fields in JSON Out](<https://devfeed.tech/articles/dharma-field-extractor-any-document-in-structured-fields-in-json-out-118022.md>)

Original publisher: [Read original article](<https://huggingface.co/blog/Dharma-AI/dharma-field-extractor>)

Author: Gabriel Pimenta de Freitas Cardoso; Paulo Henrique de Medeiros Araujo; Bruno Duarte; Francisco de Almeida Rocha Alves

Published: 2026-11-04T17:50:10Z

Content type: release

Language: en

Sources: [Hugging Face](<https://devfeed.tech/sources/hugging-face-blog.md>)

Topics: [document ai](<https://devfeed.tech/topics/document-ai.md>), [JSON](<https://devfeed.tech/topics/json.md>), [Large Language Model](<https://devfeed.tech/topics/llm.md>), [Amazon Textract](<https://devfeed.tech/topics/amazon-textract.md>), [Frontier AI](<https://devfeed.tech/topics/frontier-ai.md>), [benchmark overfitting machine learning](<https://devfeed.tech/topics/benchmark-overfitting-machine-learning.md>)

Tags: [2](<https://devfeed.tech/tags/2.md>), [2019](<https://devfeed.tech/tags/2019.md>), [2020](<https://devfeed.tech/tags/2020.md>), [contracts](<https://devfeed.tech/tags/contracts.md>), [cost](<https://devfeed.tech/tags/cost.md>), [court](<https://devfeed.tech/tags/court.md>), [field](<https://devfeed.tech/tags/field.md>), [fields](<https://devfeed.tech/tags/fields.md>), [forms](<https://devfeed.tech/tags/forms.md>), [frontier](<https://devfeed.tech/tags/frontier.md>), [ids](<https://devfeed.tech/tags/ids.md>), [json](<https://devfeed.tech/tags/json.md>), [llm](<https://devfeed.tech/tags/llm.md>), [model](<https://devfeed.tech/tags/model.md>), [ocr](<https://devfeed.tech/tags/ocr.md>), [parameter](<https://devfeed.tech/tags/parameter.md>), [range](<https://devfeed.tech/tags/range.md>), [structured](<https://devfeed.tech/tags/structured.md>), [type](<https://devfeed.tech/tags/type.md>), [wide](<https://devfeed.tech/tags/wide.md>)

### AI overview

Dharma Field Extractor converts document images into typed JSON, either by finding fields without a schema or filling an exact user-provided schema; missing requested values are returned as null. The article says one model handles multiple document types without per-type fine-tuning. In the authors' 546-document, three-scenario benchmark, it scored 0.834, above four tested frontier configurations (0.741-0.816). They report costs 6-147 times lower, but compare API list-price estimates with self-hosted GPU time on one L40S, not a sales price; deployment cost depends on hardware, volume, and latency. SROIE and FUNSD were held out from training. These are the article's reported results and do not establish performance beyond its benchmark.

### Source excerpt

*A 2-billion-parameter model that extracts exactly the fields you ask for, from any document type, outperforming every frontier LLM we tested, at a fraction of the cost.* In April, we released DharmaOCR-LITE, a model specialized for OCR across a wide range of documents, from contracts and court filings to invoices, IDs and forms.

## Create AI-Ready Enterprise Content with the Syncfusion .NET Document Chunking Library \[Webinar Show Notes\]

DevFeed: [Create AI-Ready Enterprise Content with the Syncfusion .NET Document Chunking Library \[Webinar Show Notes\]](<https://devfeed.tech/articles/create-ai-ready-enterprise-content-with-the-syncfusion-net-document-chunking-library-webinar-show-notes-115480.md>)

Original publisher: [Read original article](<https://www.syncfusion.com/blogs/post/dotnet-document-chunking-ai-rag-workflows>)

Author: Anantha Lakshmi Kannan

Published: 2026-11-03T14:00:00Z

Content type: article

Language: en

Sources: [Syncfusion](<https://devfeed.tech/sources/syncfusion-blogs.md>)

Topics: [document ai](<https://devfeed.tech/topics/document-ai.md>), [.NET](<https://devfeed.tech/topics/net.md>), [AI for interactive applications](<https://devfeed.tech/topics/ai-for-interactive-applications.md>), [ASP.NET Core](<https://devfeed.tech/topics/asp-net-core.md>), [ASP.NET](<https://devfeed.tech/topics/aspnet.md>), [pdf](<https://devfeed.tech/topics/pdf.md>), [C#](<https://devfeed.tech/topics/csharp.md>)

Tags: [ai](<https://devfeed.tech/tags/ai.md>), [ai-document-processing-in-net](<https://devfeed.tech/tags/ai-document-processing-in-net.md>), [ai-document-processing-syncfusion](<https://devfeed.tech/tags/ai-document-processing-syncfusion.md>), [ai-ready-enterprise-documents](<https://devfeed.tech/tags/ai-ready-enterprise-documents.md>), [ai-workflows](<https://devfeed.tech/tags/ai-workflows.md>), [angular](<https://devfeed.tech/tags/angular.md>), [applications](<https://devfeed.tech/tags/applications.md>), [asp-net-core](<https://devfeed.tech/tags/asp-net-core.md>), [document-chunking](<https://devfeed.tech/tags/document-chunking.md>), [document-chunking-for-rag](<https://devfeed.tech/tags/document-chunking-for-rag.md>), [document-chunking-in-c-sharp](<https://devfeed.tech/tags/document-chunking-in-c-sharp.md>), [document-chunking-library-for-net](<https://devfeed.tech/tags/document-chunking-library-for-net.md>), [document-processing](<https://devfeed.tech/tags/document-processing.md>), [document-q-a-using-rag](<https://devfeed.tech/tags/document-q-a-using-rag.md>), [docx](<https://devfeed.tech/tags/docx.md>), [embedding](<https://devfeed.tech/tags/embedding.md>), [embeddings](<https://devfeed.tech/tags/embeddings.md>), [enterprise](<https://devfeed.tech/tags/enterprise.md>), [excel](<https://devfeed.tech/tags/excel.md>), [json](<https://devfeed.tech/tags/json.md>), [markdown](<https://devfeed.tech/tags/markdown.md>), [net-document-chunking](<https://devfeed.tech/tags/net-document-chunking.md>), [net-document-chunking-library](<https://devfeed.tech/tags/net-document-chunking-library.md>), [net-document-processing](<https://devfeed.tech/tags/net-document-processing.md>), [pdf](<https://devfeed.tech/tags/pdf.md>), [powerpoint](<https://devfeed.tech/tags/powerpoint.md>), [q-a](<https://devfeed.tech/tags/q-a.md>), [rag](<https://devfeed.tech/tags/rag.md>), [rag-implementation-in-asp-net-core](<https://devfeed.tech/tags/rag-implementation-in-asp-net-core.md>), [syncfusion](<https://devfeed.tech/tags/syncfusion.md>), [webinar](<https://devfeed.tech/tags/webinar.md>), [word](<https://devfeed.tech/tags/word.md>)

### AI overview

The Syncfusion .NET Document Chunking Library prepares Excel, Word, PDF, PowerPoint, and Markdown documents as structured chunks with configurable chunking modes, size, and token overlap. Its JSON output includes metadata and citation information. In the demonstrated RAG workflow, the application separately generates and stores embeddings, retrieves matching chunks, and passes their context to an AI model; those steps are not handled by the library. The library has no built-in integration with the Angular DOCX Editor, but an edited document can be exported as a DOCX Blob, sent to an ASP.NET Core backend, and processed without permanent storage. Other formats such as video and general XML are unsupported and require separate extraction or conversion. The library also does not provide syntax-aware chunking for source code; code inside Word or Markdown is treated as ordinary content.

### Source excerpt

Learn how to prepare PDF, Word, Excel, PowerPoint, and Markdown documents for AI workflows using the Syncfusion .NET Document Chunking Library. Explore chunking, JSON output, and RAG-based document Q&A.

## Xiaomi-OCR-0: A 0.8B AI Model for Document Parsing and OCR

DevFeed: [Xiaomi-OCR-0: A 0.8B AI Model for Document Parsing and OCR](<https://devfeed.tech/articles/xiaomi-ocr-0-a-0-8b-ai-model-for-document-parsing-and-ocr-113701.md>)

Original publisher: [Read original article](<https://hackernoon.com/xiaomi-ocr-0-a-08b-ai-model-for-document-parsing-and-ocr?source=rss>)

Author: aimodels44

Published: 2026-11-03T05:00:45Z

Content type: article

Language: en

Sources: [HackerNoon](<https://devfeed.tech/sources/hackernoon.md>)

Topics: [document ai](<https://devfeed.tech/topics/document-ai.md>), [vlm](<https://devfeed.tech/topics/vlm.md>)

Tags: [ai](<https://devfeed.tech/tags/ai.md>), [ai-document-processing](<https://devfeed.tech/tags/ai-document-processing.md>), [ai-model](<https://devfeed.tech/tags/ai-model.md>), [alternatives](<https://devfeed.tech/tags/alternatives.md>), [apache-2-0](<https://devfeed.tech/tags/apache-2-0.md>), [artificial-intelligence](<https://devfeed.tech/tags/artificial-intelligence.md>), [benchmark](<https://devfeed.tech/tags/benchmark.md>), [benchmarks](<https://devfeed.tech/tags/benchmarks.md>), [cases](<https://devfeed.tech/tags/cases.md>), [checkpoint](<https://devfeed.tech/tags/checkpoint.md>), [community](<https://devfeed.tech/tags/community.md>), [customer-success](<https://devfeed.tech/tags/customer-success.md>), [handwriting](<https://devfeed.tech/tags/handwriting.md>), [hardware](<https://devfeed.tech/tags/hardware.md>), [inference](<https://devfeed.tech/tags/inference.md>), [learning](<https://devfeed.tech/tags/learning.md>), [machine-learning](<https://devfeed.tech/tags/machine-learning.md>), [model](<https://devfeed.tech/tags/model.md>), [ocr](<https://devfeed.tech/tags/ocr.md>), [optical-character-recognition](<https://devfeed.tech/tags/optical-character-recognition.md>), [parsing](<https://devfeed.tech/tags/parsing.md>), [specifications](<https://devfeed.tech/tags/specifications.md>), [tasks](<https://devfeed.tech/tags/tasks.md>), [technical](<https://devfeed.tech/tags/technical.md>), [use-cases](<https://devfeed.tech/tags/use-cases.md>), [xiaomi](<https://devfeed.tech/tags/xiaomi.md>)

### AI overview

Xiaomi-OCR-0 is a 0.8B vision-language model for document parsing and OCR-centered understanding. It supports text, tables, formulas, visual question answering over text-rich images, and structured field extraction; the source also describes use on scanned, warped, photographed, skewed, and recaptured documents. Its training combines Q-Mask text anchoring, OCR-focused continued pretraining, and mixed-task reinforcement learning on an approximately 170M-sample corpus. Reported scores include 96.83 on OmniDocBench v1.6, 97.84 on UniMER-Test for formulas, and 92.27 on PubTabNet for tables. On five OCR-centric VQA benchmarks, its reported mean is 83.2; Qwen3.5-4B scores 84.9, while smaller listed models score less. These are source-reported results, not guarantees across deployments. The excerpt does not specify image limits, languages, hardware needs, or a complete inference example. Community GGUF and NInfer conversions are separate from the original checkpoint; maintainers have not evaluated the GGUF conversions.

### Source excerpt

Learn how Xiaomi-OCR-0 handles document parsing and OCR tasks, including its technical specifications, benchmark scores, use cases, and limitations.

## Unlocking Criminal Court Decisions: Flexible Information Extraction for Empirical Legal Research

DevFeed: [Unlocking Criminal Court Decisions: Flexible Information Extraction for Empirical Legal Research](<https://devfeed.tech/articles/unlocking-criminal-court-decisions-flexible-information-extraction-for-empirical-legal-research-115413.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.03051>)

Author: Tom\\'a\\v{s} Knap, Ivana Kvapil\\'ikov\\'a, Jan \\v{C}ern\\'y, Kl\\'ara Bendov\\'a, Jarom\\'ir \\v{S}avelka, Vojt\\v{e}ch Pour, Jakub Dr\\'apal

Published: 2026-11-03T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>)

Topics: [Large Language Model](<https://devfeed.tech/topics/llm.md>), [document ai](<https://devfeed.tech/topics/document-ai.md>), [human review](<https://devfeed.tech/topics/human-review.md>), [test](<https://devfeed.tech/topics/test.md>)

Tags: [criminal](<https://devfeed.tech/tags/criminal.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-cl](<https://devfeed.tech/tags/cs-cl.md>), [legal](<https://devfeed.tech/tags/legal.md>), [legal-research](<https://devfeed.tech/tags/legal-research.md>), [research](<https://devfeed.tech/tags/research.md>)

### AI overview

The paper presents an LLM pipeline for structuring criminal verdicts through attribute exploration, text-span extraction, and value normalization, with annotation instructions that legal experts can refine. In DUI cases, revising instructions was associated with extraction accuracy rising from 52% to 70% and greater agreement between human annotators. Repeated runs of the same model agreed more with one another than with gold annotations, so consistency did not guarantee correctness; disagreements among annotators and models also exposed unclear instructions. For driving despite a ban, instructions specifying fixed categories and output formats yielded high agreement for most attributes.

### Source excerpt

arXiv:2611.03051v1 Announce Type: cross Abstract: Criminal verdicts contain descriptions of criminal behavior that can support quantitative legal research, but structuring them requires attributes and extraction rules that cannot always be fully specified in advance. We present an LLM-based pipeline that builds a structured dataset through attribute exploration, text span extraction, and value normalization, guided by annotation instructions that legal experts can refine. We evaluate it on two traffic offenses: driving under the influence (DUI) and driving despite a ban (DDB). For DUI, revising the instructions was associated with higher extraction accuracy (from 52% to 70% of correctly extracted values) and higher agreement between human annotators. Repeated runs of the same model agreed with each other more than with the gold annotations, showing that consistent output is not necessarily correct, while disagreement between annotators and between models helped reveal unclear instructions. For DDB, where the instructions defined fixed categories and output formats, agreement was high for most attributes.

## Pharos-ESG: A Framework for Multimodal Parsing, Contextual Narration, and Hierarchical Labeling of ESG Report

DevFeed: [Pharos-ESG: A Framework for Multimodal Parsing, Contextual Narration, and Hierarchical Labeling of ESG Report](<https://devfeed.tech/articles/pharos-esg-a-framework-for-multimodal-parsing-contextual-narration-and-hierarchical-labeling-of-esg-report-107846.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2511.16417>)

Author: Yan Chen, Yu Zou, Jialei Zeng, Haoran You, Xiaorui Zhou, Aixi Zhong

Published: 2026-10-30T04:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>)

Topics: [document ai](<https://devfeed.tech/topics/document-ai.md>), [Artificial Intelligence](<https://devfeed.tech/topics/ai.md>), [Data analysis](<https://devfeed.tech/topics/data-analysis.md>)

Tags: [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [esg](<https://devfeed.tech/tags/esg.md>), [framework](<https://devfeed.tech/tags/framework.md>), [labeling](<https://devfeed.tech/tags/labeling.md>), [multimodal](<https://devfeed.tech/tags/multimodal.md>), [narration](<https://devfeed.tech/tags/narration.md>), [parsing](<https://devfeed.tech/tags/parsing.md>), [report](<https://devfeed.tech/tags/report.md>)

### AI overview

Pharos-ESG converts ESG reports into structured representations to address disorderly reading order from irregular layouts and implicit hierarchies in lengthy content. It models reading order using layout flow, segments content with table-of-contents anchors, and turns visual elements into contextual natural-language descriptions through multimodal aggregation. Outputs include ESG, GRI, and sentiment labels. The abstract reports that it outperforms dedicated document-parsing systems and general-purpose multimodal models on annotated benchmarks, without providing scores or experimental details. The authors also introduce Aurora-ESG, a public dataset spanning Mainland China, Hong Kong, and U.S. markets, with structured multimodal content and fine-grained layout and semantic annotations.

### Source excerpt

arXiv:2511.16417v4 Announce Type: replace Abstract: Environmental, Social, and Governance (ESG) principles are reshaping the foundations of global financial governance, transforming capital allocation architectures, regulatory frameworks, and systemic risk coordination mechanisms. However, as the core medium for assessing corporate ESG performance, the ESG reports present significant challenges for large-scale understanding, due to chaotic reading order from slide-like irregular layouts and implicit hierarchies arising from lengthy, weakly structured content. To address these challenges, we propose Pharos-ESG, a unified framework that transforms ESG reports into structured representations through multimodal parsing, contextual narration, and hierarchical labeling. It integrates a reading-order modeling module based on layout flow, hierarchy-aware segmentation guided by table-of-contents anchors, and a multi-modal aggregation pipeline that contextually transforms visual elements into coherent natural language. The framework further enriches its outputs with ESG, GRI, and sentiment labels, yielding annotations aligned with the analytical demands of financial research. Extensive experiments on annotated benchmarks demonstrate that Pharos-ESG consistently outperforms both dedicated document parsing systems and general-purpose multimodal models. In addition, we release Aurora-ESG, the first large-scale public dataset of ESG reports, spanning Mainland China, Hong Kong, and U.S. markets, featuring unified structured representations of multi-modal content, enriched with fine-grained layout and semantic annotations to better support ESG integration in financial governance and decision-making.

## ARBITRA: Robust Document-Level ATT&CK Extraction via Multi-Expert Argumentation

DevFeed: [ARBITRA: Robust Document-Level ATT&CK Extraction via Multi-Expert Argumentation](<https://devfeed.tech/articles/arbitra-robust-document-level-att-ck-extraction-via-multi-expert-argumentation-109071.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2610.39433>)

Author: Youngjin Jin, Seungwon Shin

Published: 2026-10-30T04:00:00Z

Content type: article

Language: en

Sources: [arXiv: Cryptography and Security](<https://devfeed.tech/sources/arxiv-cryptography-and-security.md>)

Topics: [llm-reasoning](<https://devfeed.tech/topics/llm-reasoning.md>), [Large Language Model](<https://devfeed.tech/topics/llm.md>), [document ai](<https://devfeed.tech/topics/document-ai.md>), [LLM security](<https://devfeed.tech/topics/llm-security.md>), [Parallel reasoning](<https://devfeed.tech/topics/parallel-reasoning.md>)

Tags: [adversarial](<https://devfeed.tech/tags/adversarial.md>), [behavior](<https://devfeed.tech/tags/behavior.md>), [candidate](<https://devfeed.tech/tags/candidate.md>), [cs-cl](<https://devfeed.tech/tags/cs-cl.md>), [cs-cr](<https://devfeed.tech/tags/cs-cr.md>), [cyber-threat-intelligence](<https://devfeed.tech/tags/cyber-threat-intelligence.md>), [experts](<https://devfeed.tech/tags/experts.md>), [fine-tuning](<https://devfeed.tech/tags/fine-tuning.md>), [framework](<https://devfeed.tech/tags/framework.md>), [frontier](<https://devfeed.tech/tags/frontier.md>), [frontier-models](<https://devfeed.tech/tags/frontier-models.md>), [full](<https://devfeed.tech/tags/full.md>), [gpt](<https://devfeed.tech/tags/gpt.md>), [gpt-5-5](<https://devfeed.tech/tags/gpt-5-5.md>), [llm](<https://devfeed.tech/tags/llm.md>), [llms](<https://devfeed.tech/tags/llms.md>), [mappings](<https://devfeed.tech/tags/mappings.md>), [mitre](<https://devfeed.tech/tags/mitre.md>), [multiple](<https://devfeed.tech/tags/multiple.md>), [natural-language](<https://devfeed.tech/tags/natural-language.md>), [performance](<https://devfeed.tech/tags/performance.md>), [robust](<https://devfeed.tech/tags/robust.md>)

### AI overview

ARBITRA extracts MITRE ATT&CK entities from full cyber threat intelligence reports by having role-specialized LLM experts propose candidates, then checking each through a quantitative bipolar argumentation framework grounded in report evidence and rationales. The authors also build reasoning-specialized corpora to fine-tune the experts. They report a more precision-oriented operating point than sentence-level pipelines, other LLM extraction systems, and frontier models including GPT-5.5. In tests with conflicting TTP statements, cross-report contamination, and semantic poisoning, ARBITRA's mean precision degradation was roughly half that of frontier-model baselines. The supplied text is an abstract and does not provide detailed benchmark results.

### Source excerpt

arXiv:2610.39433v1 Announce Type: cross Abstract: Cyber threat intelligence (CTI) reports describe adversary behavior in natural language, but converting them into structured MITRE ATT&CK mappings remains difficult: existing methods decompose reports to sentence-level granularity for extraction, while modern LLMs can work with full reports in a single pass but with noticeable performance degradation. We present ARBITRA, which orchestrates multiple role-specialized LLM experts that propose candidate ATT&CK entities from the full report, which are then verified through a per-candidate quantitative bipolar argumentation framework (QBAF) grounded in document evidence and rationales. We additionally construct reasoning-specialized corpora for fine-tuning so experts can act as effective participants during this verification process. Across ATT&CK extraction benchmarks, ARBITRA reaches a stronger precision-oriented operating point than sentence-level pipelines, recent LLM-based extraction systems, and frontier models including GPT-5.5. Under adversarial perturbations simulating conflicting TTP statements, cross-report contamination, and semantic poisoning, ARBITRA limits mean precision degradation to roughly half that of frontier-model baselines for a more robust extraction.

## Which OCR Solution Works Best for Your Documents?

DevFeed: [Which OCR Solution Works Best for Your Documents?](<https://devfeed.tech/articles/which-ocr-solution-works-best-for-your-documents-105464.md>)

Original publisher: [Read original article](<https://blog.dailydoseofds.com/p/which-ocr-solution-works-best-for>)

Author: Avi Chawla

Published: 2026-10-27T23:58:18Z

Content type: comparison

Language: en

Sources: [Daily Dose of Data Science](<https://devfeed.tech/sources/daily-dose-of-data-science.md>)

Topics: [document ai](<https://devfeed.tech/topics/document-ai.md>), [llm-reasoning](<https://devfeed.tech/topics/llm-reasoning.md>), [llama.cpp](<https://devfeed.tech/topics/llama-cpp.md>)

Tags: [2](<https://devfeed.tech/tags/2.md>), [22](<https://devfeed.tech/tags/22.md>), [3](<https://devfeed.tech/tags/3.md>), [about](<https://devfeed.tech/tags/about.md>), [adapter](<https://devfeed.tech/tags/adapter.md>), [agent](<https://devfeed.tech/tags/agent.md>), [agentic](<https://devfeed.tech/tags/agentic.md>), [agentic-systems](<https://devfeed.tech/tags/agentic-systems.md>), [agents](<https://devfeed.tech/tags/agents.md>), [align](<https://devfeed.tech/tags/align.md>), [all](<https://devfeed.tech/tags/all.md>), [also](<https://devfeed.tech/tags/also.md>), [anthropic](<https://devfeed.tech/tags/anthropic.md>), [best](<https://devfeed.tech/tags/best.md>), [deepseek](<https://devfeed.tech/tags/deepseek.md>), [ocr](<https://devfeed.tech/tags/ocr.md>), [pages](<https://devfeed.tech/tags/pages.md>), [solution](<https://devfeed.tech/tags/solution.md>), [vs](<https://devfeed.tech/tags/vs.md>), [works](<https://devfeed.tech/tags/works.md>)

### AI overview

On a 150-page comparison, Datalab scored highest overall (88.2), but DeepSeek-OCR led on tables (87.2), and Datalab passed 5 of 6 handwriting tests while the other parsers passed none. On a filled 1040 form, PaddleOCR-VL returned only the title; Datalab passed 8 of 8 tests and DeepSeek-OCR 6 of 8. The results show that average scores can hide document-specific failures, so parser choice depends on the types of pages being processed. The open-source OmniParse Bench evaluates text, tables, and layout with pass/fail tests, records failure reasons, and lets users filter scores by traits such as handwriting, rotation, language, and script. Datalab created the benchmark and also makes Surya, so its open scorer and rerunnable tests provide a way to inspect the results directly. The excerpt recommends testing on document-matched slices and reviewing failures.

### Source excerpt

...DeepSeek-OCR vs. PaddleOCR-VL vs. Datalab Surya, tested on the same 150 pages.

## From spot fixes to measured gains: improving document extraction accuracy with ai\_extract and MLflow Prompt Registry

DevFeed: [From spot fixes to measured gains: improving document extraction accuracy with ai\_extract and MLflow Prompt Registry](<https://devfeed.tech/articles/from-spot-fixes-to-measured-gains-improving-document-extraction-accuracy-with-ai-extract-and-mlflow-prompt-registry-102795.md>)

Original publisher: [Read original article](<https://www.databricks.com/blog/spot-fixes-measured-gains-improving-document-extraction-accuracy-aiextract-and-mlflow-prompt>)

Author: Kristen Drexinger; Seshendranath Balla

Published: 2026-10-27T18:00:00Z

Content type: article

Language: en

Sources: [Databricks](<https://devfeed.tech/sources/databricks.md>)

Topics: [document ai](<https://devfeed.tech/topics/document-ai.md>), [Enhancing AI model performance](<https://devfeed.tech/topics/enhancing-ai-model-performance.md>), [AI for data automation](<https://devfeed.tech/topics/ai-for-data-automation.md>), [Data Quality](<https://devfeed.tech/topics/data-quality.md>), [Machine learning](<https://devfeed.tech/topics/machine-learning.md>), [prompt](<https://devfeed.tech/topics/prompt.md>), [Data Science](<https://devfeed.tech/topics/data-science.md>)

Tags: [ai](<https://devfeed.tech/tags/ai.md>), [best-practices](<https://devfeed.tech/tags/best-practices.md>), [customers](<https://devfeed.tech/tags/customers.md>), [data-science-and-ml](<https://devfeed.tech/tags/data-science-and-ml.md>), [document-processing](<https://devfeed.tech/tags/document-processing.md>), [engineering](<https://devfeed.tech/tags/engineering.md>), [media-entertainment](<https://devfeed.tech/tags/media-entertainment.md>), [mlflow](<https://devfeed.tech/tags/mlflow.md>), [prompt](<https://devfeed.tech/tags/prompt.md>), [registry](<https://devfeed.tech/tags/registry.md>), [solutions](<https://devfeed.tech/tags/solutions.md>)

### AI overview

Reliable document extraction requires evaluating changes across the full field set: a schema tweak that cut revision errors roughly in half increased total errors by 9%. The described workflow puts field-specific rules in schema descriptions, document-wide guidance in global instructions, uses deterministic code for normalization, and applies ai_extract precision mode when complex tables or repeated structures are the bottleneck. Citations help diagnose whether a wrong value came from extraction or reviewer-label disagreement; confidence scores can prioritize review but are not an automatic quality gate. The workflow versions both prompts and schemas in MLflow Prompt Registry, compares production and candidate configurations on the same fixed, reviewer-approved test set, and tracks accuracy, coverage, and run-to-run variation. The article reports that moving field rules into schema descriptions reduced errors 16%; precision mode alone reduced them 24%; combined, errors fell about 32% and revision errors nearly 64%. Promotion remains human-controlled, with severe field-level regressions able to block a candidate even when aggregate results look acceptable.

### Source excerpt

Document extraction often starts with a simple loop: spot an incorrect field, adjust...

## Baidu Open-Sources Unlimited OCR 2 for Hundred-Page Document Parsing

DevFeed: [Baidu Open-Sources Unlimited OCR 2 for Hundred-Page Document Parsing](<https://devfeed.tech/articles/baidu-open-sources-unlimited-ocr-2-for-hundred-page-document-parsing-101560.md>)

Original publisher: [Read original article](<https://www.marktechpost.com/2026/10/26/baidu-unlimited-ocr-2-open-source-long-document-ocr/>)

Author: Asif Razzaq

Published: 2026-10-27T05:26:36Z

Content type: article

Language: en

Sources: [MarkTechPost](<https://devfeed.tech/sources/marktechpost.md>)

Topics: [document ai](<https://devfeed.tech/topics/document-ai.md>), [Open Source](<https://devfeed.tech/topics/open-source.md>)

Tags: [ai-shorts](<https://devfeed.tech/tags/ai-shorts.md>), [applications](<https://devfeed.tech/tags/applications.md>), [artificial-intelligence](<https://devfeed.tech/tags/artificial-intelligence.md>), [attention](<https://devfeed.tech/tags/attention.md>), [baidu](<https://devfeed.tech/tags/baidu.md>), [benchmark](<https://devfeed.tech/tags/benchmark.md>), [benchmarks](<https://devfeed.tech/tags/benchmarks.md>), [computer-vision](<https://devfeed.tech/tags/computer-vision.md>), [context](<https://devfeed.tech/tags/context.md>), [context-window](<https://devfeed.tech/tags/context-window.md>), [cuda](<https://devfeed.tech/tags/cuda.md>), [decoding](<https://devfeed.tech/tags/decoding.md>), [editors-pick](<https://devfeed.tech/tags/editors-pick.md>), [flat](<https://devfeed.tech/tags/flat.md>), [language-model](<https://devfeed.tech/tags/language-model.md>), [mit](<https://devfeed.tech/tags/mit.md>), [model](<https://devfeed.tech/tags/model.md>), [new-releases](<https://devfeed.tech/tags/new-releases.md>), [ocr](<https://devfeed.tech/tags/ocr.md>), [open](<https://devfeed.tech/tags/open.md>), [open-source](<https://devfeed.tech/tags/open-source.md>), [parsing](<https://devfeed.tech/tags/parsing.md>), [request](<https://devfeed.tech/tags/request.md>), [staff](<https://devfeed.tech/tags/staff.md>), [tech-news](<https://devfeed.tech/tags/tech-news.md>), [technology](<https://devfeed.tech/tags/technology.md>)

### AI overview

Baidu's 3.34B-parameter, MIT-licensed Unlimited OCR 2 parses ordered page images into structured text, tables, formulas, figures, headings and reading order. Its SR-SWA attention uses a fixed 512-token reference set, a 128-token local window and top-16 selection over 32-token blocks to keep per-step decoding cost nearly flat as documents grow; image encoding and total runtime still increase with page count. Baidu reports scores of 84.6 on MPDocBench and 77.1 on MarathonDocBench, ahead of other models in its tests, though the benchmark figures are self-reported and the article says MarathonDocBench has no public release. Heading hierarchy is a weaker area: 48.2 on MarathonDocBench. The model card lists Transformers for quick tests and SGLang for serving; multi-page inference uses one request to retain cross-page context. Setup requires `trust_remote_code=True`; Baidu's serving example uses a customized SGLang wheel and warns that smaller page windows weaken cross-page merging.

### Source excerpt

Baidu has released Unlimited OCR 2, an MIT-licensed 3B OCR model that parses hundred-page documents in one request. Its SR-SWA attention keeps per-step decoding cost nearly flat, scoring 77.1 on MarathonDocBench The post Baidu Open-Sources Unlimited OCR 2 for Hundred-Page Document Parsing appeared first on MarkTechPost.

## RxnExBench: Benchmarking Vision-Language Models for Structured Chemical Reaction Extraction from Scientific Literature

DevFeed: [RxnExBench: Benchmarking Vision-Language Models for Structured Chemical Reaction Extraction from Scientific Literature](<https://devfeed.tech/articles/rxnexbench-benchmarking-vision-language-models-for-structured-chemical-reaction-extraction-from-scientific-literature-99972.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2610.32221>)

Author: Jiaxing Sun, Junyuan Gao, Ruikang Zhang, Hao Zheng, Lingli Ge, Xiao An, Jiang Wu, Huazheng Zeng, Jiahe Song, Chen Zhu, Haote Yang, Jingchao Wang, Lu Chen, Yuanyuan Cao, Ruijie Zhang, Shengxin Lu, Weij

Published: 2026-10-26T04:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>)

Topics: [Artificial Intelligence](<https://devfeed.tech/topics/ai.md>), [smolvlm](<https://devfeed.tech/topics/smolvlm.md>), [llm-reasoning](<https://devfeed.tech/topics/llm-reasoning.md>), [document ai](<https://devfeed.tech/topics/document-ai.md>)

Tags: [ai](<https://devfeed.tech/tags/ai.md>), [annotation](<https://devfeed.tech/tags/annotation.md>), [arxiv](<https://devfeed.tech/tags/arxiv.md>), [automated](<https://devfeed.tech/tags/automated.md>), [benchmark](<https://devfeed.tech/tags/benchmark.md>), [benchmarking](<https://devfeed.tech/tags/benchmarking.md>), [chemistry](<https://devfeed.tech/tags/chemistry.md>), [code](<https://devfeed.tech/tags/code.md>), [complexity](<https://devfeed.tech/tags/complexity.md>), [condition](<https://devfeed.tech/tags/condition.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [data](<https://devfeed.tech/tags/data.md>), [dataset](<https://devfeed.tech/tags/dataset.md>), [evaluation](<https://devfeed.tech/tags/evaluation.md>), [fields](<https://devfeed.tech/tags/fields.md>), [impact](<https://devfeed.tech/tags/impact.md>), [information](<https://devfeed.tech/tags/information.md>), [language-models](<https://devfeed.tech/tags/language-models.md>), [reaction](<https://devfeed.tech/tags/reaction.md>), [structured](<https://devfeed.tech/tags/structured.md>), [vision-language-models](<https://devfeed.tech/tags/vision-language-models.md>)

### AI overview

RxnExBench is a benchmark for extracting structured reaction data from information-dense chemistry figures such as screen tables and substrate scopes. It contains 3,565 expert-annotated reactions and 2,361 molecular structures, organized around 15 reaction roles and 26 core fields. Its difficulty taxonomy and Oracle-MolRec protocol help separate molecular recognition from higher-level reasoning. Evaluation of nine vision-language models found that performance falls as schema complexity and reasoning demands increase; even with molecular recognition handled accurately, multimodal integration and chemistry-aware reasoning remain major bottlenecks. The abstract says the data and code will be released to support further research.

### Source excerpt

arXiv:2610.32221v1 Announce Type: new Abstract: Screen Tables and Substrate Scopes are information-dense figures widely used in synthetic chemistry literature, yet automated extraction of structured reaction data from them remains challenging. We introduce RxnExBench, a multimodal benchmark for fine-grained reaction-level evaluation. Built from high-impact chemistry journals, it contains 3,565 expert-annotated reactions and 2,361 molecular structures. The dataset adopts a unified schema with 15 reaction roles and 26 core fields, supporting downstream AI-for-Chemistry tasks such as retrosynthetic planning, yield prediction, and reaction condition recommendation. RxnExBench further provides a reasoning-driven difficulty taxonomy, a modular annotation pipeline, and the Oracle-MolRec protocol, which disentangles molecular recognition from higher-level reasoning. Evaluation of 9 state-of-the-art VLMs shows that performance degrades substantially as schema complexity and reasoning demands increase. Under the Oracle-MolRec setting, results further reveal that accurate molecular recognition alone is insufficient, with multimodal integration and chemistry-aware reasoning remaining the primary bottlenecks. We release all data and code to facilitate future research in chemistry-aware multimodal understanding and AI-for-Chemistry dataset construction. The dataset and code will be released at https://github.com/opendatalab/RxnExBench.

## How Interfaze serves 50 million requests a month on Vercel

DevFeed: [How Interfaze serves 50 million requests a month on Vercel](<https://devfeed.tech/articles/how-interfaze-serves-50-million-requests-a-month-on-vercel-97698.md>)

Original publisher: [Read original article](<https://vercel.com/blog/how-interfaze-serves-50-million-requests-a-month-on-vercel>)

Author: Kevin Sundstrom

Published: 2026-10-23T04:00:00Z

Content type: article

Language: en

Sources: [Vercel News](<https://devfeed.tech/sources/vercel-news.md>)

Topics: [Vercel](<https://devfeed.tech/topics/vercel.md>), [Serverless](<https://devfeed.tech/topics/serverless.md>), [Back end](<https://devfeed.tech/topics/backend.md>), [AI Infrastructure](<https://devfeed.tech/topics/ai-infrastructure.md>), [CI/CD](<https://devfeed.tech/topics/cicd.md>), [Large Language Model](<https://devfeed.tech/topics/llm.md>), [document ai](<https://devfeed.tech/topics/document-ai.md>)

Tags: [ai](<https://devfeed.tech/tags/ai.md>), [ai-gateway](<https://devfeed.tech/tags/ai-gateway.md>), [api](<https://devfeed.tech/tags/api.md>), [backend](<https://devfeed.tech/tags/backend.md>), [classification](<https://devfeed.tech/tags/classification.md>), [data-preprocessing](<https://devfeed.tech/tags/data-preprocessing.md>), [fallback](<https://devfeed.tech/tags/fallback.md>), [fluid](<https://devfeed.tech/tags/fluid.md>), [functions](<https://devfeed.tech/tags/functions.md>), [object-detection](<https://devfeed.tech/tags/object-detection.md>), [ocr](<https://devfeed.tech/tags/ocr.md>), [routing](<https://devfeed.tech/tags/routing.md>), [sandbox](<https://devfeed.tech/tags/sandbox.md>), [scalability](<https://devfeed.tech/tags/scalability.md>), [scraping](<https://devfeed.tech/tags/scraping.md>), [speech-to-text](<https://devfeed.tech/tags/speech-to-text.md>), [third-party](<https://devfeed.tech/tags/third-party.md>), [vercel](<https://devfeed.tech/tags/vercel.md>)

### AI overview

Interfaze's models target tasks such as document processing, OCR, scraping, speech recognition, and classification where results need to be verifiable. For example, invoice extraction returns a value with its page location and a confidence score. The company says its API serves 50 million requests a month, with its frontend and backend running as serverless functions in one codebase; Vercel Sandbox runs generated code in isolation. Fluid compute scales with traffic and supports Puppeteer scraping within Vercel's memory and timeout limits. For data cleaning, AI Gateway automatically routes requests to another provider when one fails or throttles, and lets Interfaze compare providers' performance even when they claim to host the same open source model. Interfaze also uses Workflow SDK for usage-based email sequences, with automatic retries and workflows that can resume months later with state intact. Preview deployments give model versions distinct URLs and logs, and let customers test a version before release; the Vercel portion of a release reportedly builds in about three minutes, while GPU weight hosting takes longer.

### Source excerpt

Interfaze on Vercel 50 million requests a month Backend running on Fluid compute Multiple model providers behind one fallback layer with AI Gateway Interfaze builds models for deterministic work that general-purpose LLMs handle poorly: document processing, OCR, object detection, web scraping, speech to text, and classification. As founder Yoeven Khemlani puts it, "Most models aren't deterministic by nature. That's fine when you're drafting an essay. But when a hospital or a legal team needs to process documents, you don't reach for an LLM." Ask an LLM for the total on an invoice and you get a number with no easy way to check where it came from. Interfaze's models return the number along with a bounding box showing where on the page it came from and a confidence score for how sure the model is. Every result can be checked, and the models still take prompts the way an LLM does. Everything around Interfaze's models runs on Vercel, including the backend. That spans the API serving 50 million requests a month, the sandbox that executes code the model writes, and the routing that keeps data preprocessing running when a third-party model provider goes down. The backend required no infrastructure work Interfaze's frontend and backend both run on Vercel. The backend handles every request to Interfaze's API, whether a customer is calling the model or uploading a file. All of it runs as serverless functions in the same codebase as the frontend. When a request needs computation, the model hands the code to Vercel Sandbox, which runs it in isolation. There was no separate infrastructure to set up. Fluid compute handles the heavy jobs Scalability was Yoeven's top requirement. The goal is serving the model to millions of users, and Fluid compute scales with traffic on its own. Heavier work runs there too. Vercel's memory and timeout limits are high enough that Interfaze's web scraping, which runs on Puppeteer, works out of the box. AI Gateway makes model routing reliable at scale

## Multilingual Entity Extraction for Data Security | Cyera

DevFeed: [Multilingual Entity Extraction for Data Security | Cyera](<https://devfeed.tech/articles/multilingual-entity-extraction-for-data-security-cyera-96448.md>)

Original publisher: [Read original article](<https://www.cyera.com/blog/one-classification-pipeline-to-extract-them-all-entity-extraction-across-multiple-languages>)

Author: Ayala Allon, Ph.D.

Published: 2026-10-22T23:04:44Z

Content type: article

Language: en

Sources: [Cyera Blogs](<https://devfeed.tech/sources/cyera-blogs.md>)

Topics: [document ai](<https://devfeed.tech/topics/document-ai.md>), [ner](<https://devfeed.tech/topics/ner.md>), [Fine-tuning](<https://devfeed.tech/topics/fine-tuning.md>), [Security](<https://devfeed.tech/topics/security.md>), [Data Quality](<https://devfeed.tech/topics/data-quality.md>)

Tags: [2](<https://devfeed.tech/tags/2.md>), [3](<https://devfeed.tech/tags/3.md>), [about](<https://devfeed.tech/tags/about.md>), [account](<https://devfeed.tech/tags/account.md>), [ai](<https://devfeed.tech/tags/ai.md>), [ai-powered](<https://devfeed.tech/tags/ai-powered.md>), [all](<https://devfeed.tech/tags/all.md>), [also](<https://devfeed.tech/tags/also.md>), [analysis](<https://devfeed.tech/tags/analysis.md>), [cyera](<https://devfeed.tech/tags/cyera.md>), [data](<https://devfeed.tech/tags/data.md>), [data-security](<https://devfeed.tech/tags/data-security.md>), [error](<https://devfeed.tech/tags/error.md>), [fine](<https://devfeed.tech/tags/fine.md>), [fine-tuning](<https://devfeed.tech/tags/fine-tuning.md>), [post](<https://devfeed.tech/tags/post.md>), [processing](<https://devfeed.tech/tags/processing.md>), [security](<https://devfeed.tech/tags/security.md>), [sensitive-data](<https://devfeed.tech/tags/sensitive-data.md>), [using](<https://devfeed.tech/tags/using.md>)

### AI overview

Cyera found that English fine-tuned entity extraction models generalized reasonably well to French, but some valid predictions were rejected when their wording did not match the source text. The team curated French validation samples, used an LLM to help establish ground truth, manually reviewed errors, and checked improvements on a held-out set. For rare accent and month-language mismatches, it refined post-processing to match normalized or translated forms while restoring the original French span. Entity types with gaps that post-processing could not reliably fix received multilingual fine-tuning. The excerpt reports high precision and recall for most tested entities but gives no numerical scores.

### Source excerpt

See how Cyera extended entity extraction to French using error analysis, refined post-processing, and multilingual fine-tuning to classify sensitive data.

[Next page](<https://devfeed.tech/topics/document-ai.md?cursor=WyIyMDI2LTEwLTIyVDIzOjA0OjQ0KzAwOjAwIiwgImFmNjYzMDQxLWVmMDYtNGFjMi05NTUwLWE4MjI5MDBkNGVjMyJd>)