# Turn documents into Markdown or chunks for retrieval

> Turn documents into Markdown or chunks for retrieval: 2 companies fully do it, 13 with a limit, 24 carry the tag. Every listing in Document parsing & extraction with the reason and a source.

- Canonical: https://www.anchorterminal.com/niches/document-extraction.rag-ingest/
- Markdown: https://www.anchorterminal.com/niches/document-extraction.rag-ingest/index.md (~2,000 tokens)
- Slim: https://www.anchorterminal.com/niches/document-extraction.rag-ingest/index.min.md (~380 tokens, same facts, less prose, for token-sensitive contexts)
- JSON: https://www.anchorterminal.com/niches/document-extraction.rag-ingest/index.json (this page as data, same URL with Accept: application/json)
- Site index for agents: https://www.anchorterminal.com/llms.txt (the key docs and every listing in one file, at most 16 MB, about 4 million tokens: https://www.anchorterminal.com/llms-full.txt)
- API: https://www.anchorterminal.com/api/v1/index.json
- MCP: https://www.anchorterminal.com/mcp (the whole catalogue as tools, no key)
- DNS: dig +short TXT <slug>.grade.anchorterminal.com (the grade as a DNS TXT record, for a sandbox with no HTTP where names still resolve; also top., best.<category>. and <a>.vs.<b>. under the zone, and the same under grade.letme.dev)
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)
- Updated: 2026-11-15

Two or three companies' docs show a full fit.

| Companies, full fit | With a limit | Carry the tag | Share of the category's work | First-hand accounts |
| --- | --- | --- | --- | --- |
| 2 | 13 | 24 | 16 | 10 |

## What counts as done

Markdown or JSON per document in reading order, with headings, tables and figure captions kept, page numbers on every chunk, and chunks under a stated token size, from one call or one async job per document.

Measured, for our own tests: a 20-page report returns within 60 seconds.

How agents ask for it: convert these PDFs to markdown for my RAG; ingest this folder of documents into the vector store; parse this report so I can ask questions about it; chunk these files by section

## Does the job

The docs show it meets the test above, and nothing we read says otherwise.

- **Extend API + MCP (CrowdView Inc. dba Extend)**: Parse returns chunks by page, section or document with minCharacters and maxCharacters, each with a pageRange and typed blocks with boxes, tables as Markdown or HTML and figures summarised and classified, from one async run per document.
- **Reducto API + MCP (Reducto, Inc.)**: Parse returns Markdown chunks grouped by section with chunk_size (750 to 1,250 characters at 1,000), blocks with type, page and normalised bbox, figures as image_url, and usage.num_pages; acc-19 found typed blocks with boxes and photo-only pages described.

## Does the job, with a limit

It does the job, but something in the docs or a user's account will get in an agent's way. The limit is named in each row.

- **Adobe PDF Services / PDF Extract API (Adobe Inc.)**: /operation/pdftomarkdown returns Markdown with figures as base64 (getFigures), and Extract gives ordered elements with a Page each, but there is no chunking option, so chunks under a token size are the agent's job. Files cap at 400 pages (150 scanned).
- **Amazon Textract (Amazon Web Services, Inc.)**: The Layout feature returns LAYOUT blocks with page and geometry in reading order, but output is a flat JSON block list with no Markdown, no chunking and figures not exported; acc-02 ranks Textract well below LlamaParse for RAG with tables.
- **Azure Document Intelligence (Microsoft Corporation)**: prebuilt-layout with outputContentFormat=markdown returns Markdown with HTML tables, sectionHeading, pageHeader and figures with captions, and boundingRegions give pages, but there is no chunking option, so token-sized chunks are the agent's job.
- **Datalab (Endless Labs, Inc.)**: Convert returns markdown, html, json or chunks, with images and captions, paginate and page_range, and acc-13 found Marker output accurate, but the chunk size is not stated and model output changes between updates with pinning enterprise-only.
- **Docling (Docling Project (LF AI & Data Foundation))**: Exports Markdown and lossless JSON, and HybridChunker splits by tokenizer limit keeping headings and captions with page provenance; acc-03 found output good. But acc-25 saw a new default backend drop most of the text layer with no warning, and acc-29 found no built-in image storage.
- **Google Cloud Document AI (Google LLC)**: Layout Parser chunks with chunkSize and includeAncestorHeadings, each chunk with pageSpan, headers and footers, plus returnImages and table and image annotation, but tables across pages may split in two, and files over 15 pages need batchProcess through Cloud Storage.
- **LandingAI Agentic Document Extraction (LandingAI, Inc.)**: Parse returns reading-order Markdown with page breaks, HTML tables keeping merged cells, classified figures with generated descriptions and per-block page grounding, but there is no chunking option or token-sized chunks; blocks are the nearest unit.
- **LlamaParse API + MCP (LlamaIndex, Inc.)**: Parse returns Markdown per page with items, saved images and merged continued tables, and acc-01 and acc-02 rate it highly, but Parse has no token-size chunking, acc-33 saw a two-page job take over ten minutes, and acc-35 found parsed output kept with no delete.
- **Mistral OCR API (Mistral AI (RCS Paris 952 418 325))**: Returns Markdown per page with images, separate HTML tables, optional header and footer fields and page dimensions, but there is no chunking, and acc-37 saw a header column dropped at the default DPI.
- **Nanonets API + MCP (Nano Net Technologies Inc.)**: /api/v2/parse returns Markdown or HTML, with elements and bounding boxes when metadata is requested, but the OpenAPI file has no chunking option.
- **Nemotron 3.5 Parse (NVIDIA Corporation)**: Each page image becomes Markdown with HTML tables, semantic classes including captions and page headers, and boxes, but it parses one page per request, PDFs must be rendered first, and there is no chunking or document-wide reading order.
- **OpenDocRouter (LlamaIndex, Inc.)**: POST /v1/parse returns Markdown per page with a status each, and layout adds typed elements (section_header, table, caption, page_header) with boxes, but there is no chunking and no image export; pages are the only unit.
- **Unstructured API + MCP (Unstructured Technologies, Inc.)**: Transform Parse returns Markdown or elements with type, page_number, coordinates and table_html, plus failed_pages and degraded-page warnings, but the Transform API has no chunking option; acc-02 ranked it third for complex tables and figures.
- **Xiaomi-OCR-0 (Xiaomi Corporation)**: The whole-page prompt returns Markdown in reading order with tables as OTSL and formulas as LaTeX, but one image per request means PDFs go page by page, headers and footers are dropped, and there is no chunking or figure export.

- not this job: Clairic (Cloudflare, Inc.): Clairic returns typed field values, not converted document text; it takes a text state and up to four images, with no PDF input or Markdown output.
- not this job: Invofox (Invofox, Inc. (Delaware), with a registered branch in Spain): The API reference and OpenAPI file cover typed document extraction only; no Markdown or chunk output endpoint.
- not this job: Mindee API (Mindee SAS (Mindee Inc. outside the Eurozone)): Raw Text returns plain OCR text per page; there is no Markdown, table or chunk output for retrieval.
- not this job: Veryfi API + MCP (Veryfi, Inc.): Veryfi extracts typed fields from receipts, invoices, statements, cheques and tax forms; the SDK shows no Markdown, layout or chunk output.

6 listings in the category were added after this reading and aren't marked yet.

## What stops agents finishing

- install fails on the user's platform (Docling on Windows)
- figures and images dropped with no way to store and cite them
- per-page price above the user's budget for a large corpus

## Around it

Before: file-storage.agent-upload, web-scrapers.docs-site-to-knowledge. After: vector-search.ingest-sync, company-knowledge.cited-answer.

## Where this comes from

Weight 16 of 100 in Document parsing & extraction: 10 of 46 accounts (Hacker News RAG threads, Docling and Unstructured issues); the most common reason given for choosing a parser

As data: https://www.anchorterminal.com/api/v1/use-cases/document-extraction.rag-ingest.json, or the MCP tool get_use_case with id document-extraction.rag-ingest. Draft, read 2026-11-15 from public docs and first-hand accounts, not yet reviewed by a person.

---

Query this catalogue directly over MCP: https://www.anchorterminal.com/mcp (Streamable HTTP, no key). How: https://www.anchorterminal.com/agents/#mcp
