StackMap
Subscribe

olmocr vs PixelRAG

Open toolkit that linearizes messy PDFs — scans, tables, equations, handwriting — into clean ordered Markdown with a self-hosted vision-language model. Built for LLM training data and RAG ingestion. — versus — Berkeley's visual RAG: render pages and PDFs to screenshot tiles and retrieve with a VLM embedder — tables, charts and layout survive. pixelshot CLI plus a hosted 8.28M-page Wikipedia index.

The curated verdict

olmocr linearizes messy PDFs into clean ordered Markdown so a text pipeline can index them; PixelRAG argues the linearization step is the loss — embed the page image and let a VLM read the tile directly.

olmocrPixelRAG
Stars19k6.9k
Forks1.6k580
LanguagePythonPython
LicenseApache-2.0Apache-2.0
Last activity3 months ago4 days ago
Topicsrag, ocrrag, vision
Curated connections92

olmocr — the curator's take

Reach for olmocr when you have lots of messy PDFs to convert into high-quality text for a training corpus or RAG index and you have GPU to run the VLM. It is the parsing FRONT-END of a pipeline, not a RAG system itself — pair it with an indexer/retriever (LlamaIndex) and a vector store (Chroma). NOT worth it for clean, digital-native PDFs where a cheap text extractor does the job — running a vision-language model for those is overkill.

PixelRAG — the curator's take

The paper's claim — screenshots beat parsed text for RAG — matters when your answers live in visual structure: tables, charts, infographics, layout-heavy PDFs that text chunkers flatten into noise. Zero-setup on-ramp is real: a hosted 8.28M-page Wikipedia index with no API key, and the pixelbrowse Claude Code plugin gives the agent screenshot-based page reading (pixelshot via Playwright/CDP, no MCP server). NOT a drop-in for text RAG stacks: your reader must be a VLM, tile indexes cost more storage/compute than text embeddings, and retrieval quality rides on their LoRA-tuned Qwen3-VL-Embedding model. Research codebase — expect pipeline assembly, not a product.