The open-source RAG stack

From a folder of PDFs and a list of URLs to answers you can measure — seven repos in pipeline order, with the why at every step and a way out when chunk-and-embed stops working.

7 steps · every link human-reviewed · how we curate →

1. 1
   pdf-inspector★ 19kocr · rag

   Firecrawl's Rust PDF triage: classifies text-based vs scanned in ~10-50ms, extracts positioned text and clean Markdown without OCR — routing the ~54% of PDFs that never needed a model.

   Why this stepMost pipelines OCR everything. About half of PDFs are text-based and never needed a model — this classifies each one in tens of milliseconds and extracts the text locally, so only the genuinely scanned pages go on to something expensive.
2. 2
   marker★ 39kocr · rag

   Datalab's 39k-star PDF-to-Markdown/JSON converter: a layout pipeline plus an optional LLM pass for tables, forms and equations, with chunk output and form-value extraction built in.

   Why this stepThe scanned half, and anything with tables, forms or equations. A layout pipeline that runs on your own hardware, with an optional LLM pass on only the blocks that need it. Check the model-weights licence before you ship it commercially.
3. 3
   crawl4ai★ 81kweb · rag

   The 74k-star LLM-native crawler: turns any site into clean, RAG-ready Markdown — adaptive crawling, JS rendering, extraction strategies, Docker deploy. Python, Apache-2.0.

   Why this stepThe other half of your corpus lives on the web. This is the default when the deliverable is Markdown for a model rather than rows for a database — JS rendering, content filtering, deep crawl, one Docker command.
4. 4
   LlamaIndex★ 52krag · agents

   Data framework for connecting custom data sources to LLMs — ingestion, indexing, retrieval.

   Why this stepNow wire it together. Loaders for what the first three produced, chunking, indexing and query engines in one framework — so the retrieval logic lives in one place instead of in glue code.
5. 5
   Chroma★ 29krag · memory

   Open-source embedding database for building AI apps with retrieval.

   Why this stepSomewhere to put the embeddings. The boring default: embedded or served, metadata filtering, persistence, first-class LlamaIndex support. Stay here until scale forces a choice; its page lists what to reach for then.
6. 6
   PageIndex★ 36krag

   Vectorless, reasoning-based RAG — builds a hierarchical tree index from long documents so an LLM retrieves by relevance instead of similarity. No chunking, no embeddings, no vector DB.

   Why this stepThe way out when chunk-and-embed returns "similar but irrelevant" — long, structured documents like contracts, filings and manuals. No chunks, no embeddings: a tree index the model reasons over. Costs an LLM call per query, so use it where similarity search fails, not everywhere.
7. 7
   Ragas★ 16kevals · rag

   Evaluation toolkit for your RAG and agent pipelines — faithfulness, relevance, and more.

   Why this stepRetrieval quality is invisible until you measure it. Faithfulness, answer relevance, context precision and recall — run it before every change to the chunker, the embedder or the store, so you know which one moved the number.