Blog

Extract tables from PDF files in Python using pdfmux, PyMuPDF, and Docling. Code examples, accuracy benchmarks, and the best approach for each use case.

We benchmarked 8 Python PDF libraries on 200 real-world PDFs. Ranked by table accuracy, reading order, and heading structure. Updated May 2026.

Extract text, tables, and structured data from PDFs in Node.js. Library comparison (pdf-parse, unpdf, pdf2json), benchmarks, code, and the pdfmux MCP path.

PDF to JSON for LLM tool use, function calling, and structured outputs. 3-layer schema, bounding boxes, confidence scoring, and the pdfmux pattern.

How pdfmux achieves 99% of AI-powered extraction accuracy with zero GPU, zero API keys, and zero cost per page. Architecture explained.

Three reliable ways to detect a scanned PDF in Python, plus a routing pattern that runs OCR only on the pages that need it. With code and benchmarks.

Four chunking strategies for PDF-backed RAG pipelines, benchmarked on retrieval recall. With code, a comparison table, and a recommendation by document type.

Two-minute setup to give Claude Code, Claude Desktop, and Cursor any-PDF parsing via the pdfmux MCP server. Local, offline, no API keys.

Three ways to load PDFs into LlamaIndex in 2026: SimpleDirectoryReader, PDFReader, and a custom file_extractor. Plus the pdfmux integration that gives you a confidence score per page.

Honest 2026 comparison of pdfmux and AWS Textract on accuracy, cost, privacy, and integration. Where Textract still wins, and where the cost math has flipped against it.

No single PDF extraction tool wins at every category. I tested PyMuPDF, Docling, Marker, Surya, Mistral OCR, Gemma 4, and Gemini Flash across digital, table-heavy, scanned, and Arabic PDFs — then built pdfmux, a self-healing pipeline that picks the best extractor per page.

433 customer PDFs. v1: 16 silent failures. v4: zero. The CLI exit code lied to us. Here's exactly what failed and the 1.6.1+1.6.2 fix.

How to load PDFs into LangChain. PyPDFLoader, UnstructuredPDFLoader, PyMuPDFLoader, and the pdfmux loader compared on speed, table fidelity, and cost.

What shipped in pdfmux 1.6.0: Mistral OCR, Marker, Gemma 4, hash-keyed result cache, NDJSON streaming, configuration profiles, watch mode, cost prediction, diff, retry with backoff. With the exact CLI commands.

Honest head-to-head: PyMuPDF vs pdfplumber on speed, table extraction, license, and Python ergonomics. Benchmarks on 1,422 real pages plus a third option.

Process thousands of PDFs in Python: memory-safe batching, parallel workers, and error recovery. Benchmarked patterns for production PDF pipelines.

Extract form data from PDFs in Python: AcroForms, XFA, and mixed docs covered. Code for pypdf, pdfrw, pdfmux with real-world examples.

Direct, benchmark-backed comparison of pdfmux, LlamaParse, Docling, and Unstructured for RAG pipelines in 2026 — accuracy scores, pricing, install size, GPU requirements, and a per-use-case decision matrix.

How to extract invoice data from PDFs in Python. Benchmarks on 1,200 real invoices, production code for field extraction, and a working AP automation pipeline.

Head-to-head: pdfmux versus LlamaParse on accuracy, cost, and privacy. Which PDF extractor fits your RAG pipeline in 2026 — and when does the cost math flip?

Extract bilingual Arabic and English data from Bills of Lading, commercial invoices, and customs forms. Production code, benchmark data, and compliance notes for UAE and Saudi e-invoicing deadlines.

The only PDF MCP server with confidence scoring, multi-method extraction, and structured output. Install in 30 seconds with npx -y pdfmux. Works in Claude Desktop, Cursor, Windsurf, and VS Code.

Replace cloud LLM extraction with Gemma 4 running locally via Ollama. Zero API costs, full privacy, 140+ language OCR. Hardware requirements, benchmark numbers, and setup commands.

AI agents need to read PDFs to do real work — invoices, contracts, reports. But most extraction tools return text with no quality signal. Here is the pattern emerging across production agent systems, with code examples for Claude, GPT, and LangChain.

OpenDataLoader PDF v2.0 tops benchmarks at 0.90 accuracy — but production RAG pipelines need more than one extractor. How pdfmux uses OpenDataLoader alongside PyMuPDF, Docling, and OCR to handle the documents that break any single tool.

LiteLLM, the most popular open-source LLM proxy, was compromised via a poisoned security scanner in its CI/CD pipeline. Here is what happened, why it matters, and what it means for AI infrastructure security.

How to extract text from scanned and image-based PDFs in Python using OCR. Compares pytesseract, EasyOCR, Surya, and pdfmux with code examples and accuracy benchmarks.

Build a production RAG pipeline with clean PDF extraction. Covers pdfmux + LangChain + ChromaDB with code, chunking strategies, and how extraction quality directly impacts retrieval accuracy.

Head-to-head comparison of pdfmux and Kreuzberg for Python PDF extraction. Benchmark results on 200 PDFs, code examples, feature comparison, and which to use for RAG pipelines.

How to convert PDFs to clean Markdown for RAG and LLM pipelines. Covers chunking, table handling, OCR, quality scoring, and production patterns.

Head-to-head comparison of Python PDF extraction libraries on 200 real-world PDFs. Benchmark scores, code examples, pricing, and which to use for RAG pipelines.

There are now 7+ serious PDF extraction tools — OpenDataLoader, Docling, Marker, MinerU, pymupdf4llm, MarkItDown, pdfmux, and more. Here is when to use each one, with real benchmarks, cost breakdowns, and honest tradeoffs.

Most PDF extractors run once and hope for the best. pdfmux extracts, audits every page with 5 quality checks, and re-extracts failures automatically — here is exactly how the algorithm works, with real page scores and before/after examples.

I maintain a tool that uses PyMuPDF, Docling, Surya, RapidOCR, and Gemini Flash internally. Here is when each one wins, when each one fails, and how to pick the right one for your use case — with real numbers.

There are 18,000+ MCP servers but most PDF ones just wrap basic text extraction. Here is how to add production-grade PDF processing to Claude, Cursor, or any MCP-compatible AI agent — with confidence scoring, table extraction, and OCR built in.

Real-world benchmark of pdfmux vs PyMuPDF and pymupdf4llm across 11 public documents — 1,422 pages of SEC filings, academic papers, legal opinions, and government reports. Full results with every number.