Top GitHub Repositories for RAG in 2026: Ingestion, Engines, Graphs and Vectors

Short version: RAG in 2026 is no longer a pipeline you assemble from a vector database and a prompt. The projects that grew fastest this year sit at the edges: document ingestion (Markitdown, MinerU, Docling), graph construction (Graphify, LightRAG, Graphiti) and engines that call themselves context infrastructure rather than RAG (RAGFlow hit 1.0 on 6 November). Below is the map, layer by layer, with star counts, licences and latest releases read from the GitHub API on 10 November 2026, three reference stacks at the end, and the projects that were archived this year. It is refreshed monthly.

Ingestion: where most RAG quality is won or lost

Bad chunks make bad answers, and in 2026 the ingestion layer is where the stars went. microsoft/markitdown (194,673 stars, MIT, v0.1.9) converts Office files, PDFs and more to Markdown and has become the default first step for anything LLM-bound. opendatalab/MinerU (83,417 stars, 4.1.7 on 4 November) is the heavy tool for complex PDFs, with layout analysis that keeps tables and formulas intact; its README now pitches it as a command-line document reader for agents. docling-project/docling (70,356 stars, MIT) sits in between: a Python library with a strong document model that releases several times a week (v2.150.0 on 10 November). Unstructured-IO/unstructured (15,708 stars) is the ETL-style option. Our rule: Markitdown for everyday files, Docling when you need structure, MinerU when the PDFs fight back.

Web data: Firecrawl and Crawl4AI

If the knowledge lives on the web rather than in files, firecrawl/firecrawl (200,289 stars) and unclecode/crawl4ai (87,170 stars) are the two standards. They differ in licence as much as features: Firecrawl is AGPL-3.0, which matters if you ship it inside a product; its last published GitHub release is v2.11.0 from June 2026, although patch tags (v2.11.7xx) and commits land daily; Crawl4AI is Apache-2.0 with v0.9.5 from 14 October, and its README now leads with a hosted cloud API alongside the open-source crawler.

Engines: RAGFlow 1.0, Dify, AnythingLLM

infiniflow/ragflow (93,944 stars, Apache-2.0) is the most important release in this list: v1.0.0 shipped on 6 November 2026 as a complete rewrite in Go, and its GitHub description now calls it an agent-native database, a RAG engine that grew into context infrastructure. Note that it runs on Elasticsearch, Infinity or OpenSearch as its document engine, not on the general-purpose vector databases below. Its deep document understanding is still the reason to choose it. langgenius/dify (160,816 stars, 1.19.0) is the broader platform: agentic workflows plus RAG pipelines in one collaborative workspace, under Dify's own licence rather than a standard open-source one. Mintplex-Labs/anything-llm (68,132 stars, MIT, v1.19.0) is the local-first desktop and server app for people who want their documents and agents on their own machine. onyx-dot-app/onyx (33,317 stars, v5.1.1), formerly Danswer, is the enterprise chat-over-everything option, and pathwaycom/serviette (58,722 stars, MIT), formerly llm-app, is a no-code RAG server that keeps answers in sync with changing documents and works with eight vector databases.

Frameworks: LangChain, LlamaIndex, Haystack

The three orchestration frameworks are all alive and all repositioned. langchain-ai/langchain (149,295 stars, MIT) is on its 1.x line (langchain-core 1.7.7 on 10 November) and calls itself the agent engineering platform; the classic retrievers moved to the langchain-classic package, which trips up every tutorial written before October 2025. run-llama/llama_index (52,866 stars, MIT, v0.14.25) now describes itself as the document processing platform for AI, and its README says plainly that the team's primary focus has shifted to LlamaParse while the open-source framework remains available as a toolkit; plan accordingly. deepset-ai/haystack (26,975 stars, Apache-2.0, v2.31.1 on 3 November) is the most explicitly production-minded of the three and the one we would pick for a pipeline that has to run unattended. neuml/txtai (13,254 stars, v9.14.0) deserves a mention as the all-in-one library that never chased hype.

Graph RAG grew up

Two years ago graph RAG meant Microsoft's research project. In 2026 it is a category. Graphify-Labs/graphify (145,251 stars, Apache-2.0, created April 2026) turns a codebase with its docs, SQL schemas, configs and PDFs into a queryable knowledge graph, and ships as a /graphify skill for coding agents, which is why it outgrew everything else in this section. HKUDS/LightRAG (40,635 stars, MIT, EMNLP 2025) is the simple, fast graph RAG most teams start with, and its sibling RAG-Anything (24,481 stars) extends it to multimodal documents. microsoft/graphrag (36,619 stars, v3.3.0 on 8 October) remains the reference implementation, though its README now says the project is largely in maintenance mode: bug fixes and CVE updates only, no new features. getzep/graphiti (32,704 stars, Apache-2.0, v0.31.2) is the one to use when the graph has to change in real time, which is the agent memory case.

Vector stores

The vector database war is over and everyone won a niche. milvus-io/milvus (46,629 stars, v3.0.5) for very large scale, qdrant/qdrant (35,497 stars, v1.20.0, Rust) for performance with a friendly API, pgvector/pgvector (23,639 stars) when you already run Postgres and would rather not add a service, weaviate/weaviate (16,945 stars, v1.40.4) for hybrid object-plus-vector search, lancedb/lancedb (11,872 stars, v0.41.0) as the embedded, multimodal option, and chroma-core/chroma (29,667 stars, now Rust) as the quickest local start, with a cloud offering behind it. facebookresearch/faiss (41,301 stars, v1.15.1) is still the library under many of them.

Evaluation and learning

ragas (16,280 stars, now under the vibrantlabsai organisation) is still the common vocabulary for RAG metrics, though its last release was v0.4.3 in January 2026 and the repository has been quiet since February, so check it is still the right fit. NirDiamant/RAG_Techniques (30,287 stars) remains the best single place to learn the patterns, one notebook per technique.

Archived or moved in 2026

• vanna-ai/vanna (text-to-SQL RAG) archived March 2026 (last release v2.0.2, February 2026).
• microsoft/graphrag is not archived but now describes itself as largely in maintenance mode.
• weaviate/Verba archived June 2026.
• QuivrHQ/quivr moved to The-Vibe-Company/quivr and was rebuilt in Go as an engine for continuous content streams; it is no longer the "chat with your documents" app people remember.
• microsoft/kernel-memory now lives at dluc/kernel-memory, whose README calls it an archived research project.
• stanford-oval/storm: no push since September 2025.
• Cinnamon/kotaemon: still available (v0.12.0, May 2026) but slowing, last push July 2026.

Three reference stacks

• Solo, local, free: Docling or Markitdown for ingestion, LanceDB or Chroma for vectors, Haystack or LlamaIndex to glue it, an Ollama model to answer. Runs on a laptop.
• Team knowledge base: RAGFlow 1.0 (on Elasticsearch, Infinity or OpenSearch) or Dify (on Qdrant or Milvus) as the engine, MinerU for the PDF backlog, Haystack if you need custom pipelines beside the UI.
• Agents with memory: Graphiti or LightRAG for the graph, pgvector for plain retrieval, Graphify if the corpus is a codebase. Graphiti and Graphify both ship MCP servers; LightRAG and pgvector need a thin wrapper to be exposed to an agent the same way.

How this list is built

Star counts, licences, release tags and archive status come from the GitHub API, captured on 10 November 2026; positioning statements come from each project's README on the same day. We include a project when it does a job in a real RAG stack and has shipped a release or tag in 2026, and we note per entry when activity has slowed; frameworks, engines and infrastructure are grouped by layer rather than forced into one ranking. We re-run the numbers monthly and log changes at the top of this page. For the agent frameworks that consume these stacks, see our AI agent repositories list.