PDF to Markdown — Tables + OCR, for RAG & AI Agents

Convert PDFs into clean, LLM-ready markdown at scale. Every page is quality-scored and routed to the right engine: fast layout-aware extraction for born-digital pages, and a vision-model OCR tier for scanned, broken, or table-dense pages — so one bad scan doesn't wreck your pipeline, and you don't pay OCR prices for clean pages.

Unofficial independent tool. Your documents are processed only to produce your output.

Convert PDFs to markdown for RAG

Feed a list of PDF URLs, get markdown with headings, lists, and tables preserved. Use perPageRows: true to receive one row per page — ready for chunking into a vector store. Each page reports its extraction method (text or vlm) and quality signals (character count, garbage ratio, image coverage), so you always know what happened.

OCR scanned PDFs with a vision model

Pages with no usable text layer — scans, faxes, image-only exports, PDFs with broken font maps — are rendered and transcribed by a vision language model into proper markdown, tables included. Three modes via llmOcr:

• auto (default): only pages that need it are billed as ocr-page events
• off: deterministic extraction only — no LLM ever sees your document
• always: force every page through the vision model for maximum fidelity

Extract tables from PDF to markdown

Born-digital tables are extracted from the layout; ambiguous or scanned tables go through the vision tier and come back as GitHub-flavored markdown tables instead of tab-soup.

Never-fail runs

Corrupt files, password-protected PDFs, and dead URLs come back as structured error rows ({url, error}) — the run itself succeeds, your pipeline keeps moving, and you only pay for pages actually processed. User-set spending caps are respected mid-run: processing stops cleanly at your limit with charge_limit_reached: true on the row.

Output schema

Per document:

url, pages_processed, pages_total, ocr_pages, methods, vlm_usage, markdown, char_count

url, page, markdown, method, chars, alpha_ratio, garbage, image_cover

Use with AI agents

MCP-friendly: an agent can hand this actor a PDF URL and get structured markdown back in one call, paying cents per document. Pairs naturally with web-crawling actors — crawl, collect PDF links, convert here, feed your RAG store.