Most “best RAG framework” lists still rank tools by features and GitHub stars. When I compared a year of commit history instead, three things stood out that those lists miss: LlamaIndex’s open-source framework is now a side project for its company, Microsoft’s GraphRAG is in maintenance mode, and Weaviate has started putting some new features under a separate enterprise licence. None of that makes them bad choices, but it should change how you pick a tool you’ll depend on for years.
Live tracker: the numbers in this article are a snapshot from 15 October 2026. For the current status of these projects, and more than 100 open-source AI tools in total, each with a 0 to 100 health score, see the AI Tool Health Index, which refreshes from GitHub automatically. Four of these are also compared side by side: LangChain vs LlamaIndex and Qdrant vs Milvus. Two differences to expect: the tracker counts every commit, including merge commits, so projects that merge pull requests that way (Weaviate and LightRAG, for example) show higher numbers there than in the table below; and for Qdrant the tracker follows the dev branch, where day-to-day work happens, while this article counted the default master branch.
I haven’t built a RAG system with any of these tools. My own “chat with my documents” setup is just files uploaded to Claude Projects. So instead of ranking them by my taste, I looked at who is still working on each one.
Retrieval-augmented generation (RAG) stacks usually have two parts: a framework that handles loading, chunking, retrieval and prompting, and a vector database that stores embeddings. Both are long-term dependencies. So instead of repeating feature lists, I looked at the evidence of who’s still investing in each project.
How This Was Measured
Method: on 15 October 2026 we cloned the public GitHub repositories of 14 RAG frameworks and vector databases and compared two identical 91-day windows: 17 July to 15 October 2025, and 17 July to 15 October 2026. For each window we counted commits on the default branch (excluding merge commits), distinct commit-author email addresses (excluding bots such as dependabot and project release bots), and, for 2026, commits that credit an AI coding tool as co-author. We also read each repository’s licence file and README. The counting scripts were written and run with AI assistance in a cloud sandbox and the key findings were checked by hand. We did not install or benchmark any of these tools.
Commit counts are a rough signal. Projects that squash pull requests into one commit look smaller than projects that keep every commit, and a busy repository can still be the wrong tool for you. Treat big changes in direction, and what a project says about itself, as the useful part.
The Results
| Project | Type | Commits 2025 → 2026 | Authors 2025 → 2026 | Last commit | Licence |
|---|---|---|---|---|---|
| LlamaIndex | Framework | 362 → 105 | 141 → 57 | 14 Oct 2026 | MIT |
| LangChain | Framework | 739 → 552 | 141 → 57 | 15 Oct 2026 | MIT |
| Haystack | Framework | 135 → 744 | 24 → 131 | 15 Oct 2026 | Apache 2.0 |
| RAGFlow | RAG engine | 669 → 2,709 | 93 → 244 | 15 Oct 2026 | Apache 2.0 |
| LightRAG | Graph RAG | 828 → 1,476 | 35 → 64 | 14 Oct 2026 | MIT |
| GraphRAG (Microsoft) | Graph RAG | 22 → 23 | 5 → 3 | 8 Oct 2026 | MIT |
| DSPy | Framework | 160 → 185 | 41 → 45 | 15 Oct 2026 | MIT |
| txtai | Framework | 49 → 224 | 1 → 61 | 15 Oct 2026 | Apache 2.0 |
| Chroma | Vector DB | 318 → 120 | 22 → 10 | 13 Oct 2026 | Apache 2.0 |
| Qdrant | Vector DB | 284 → 970 | 21 → 55 | 5 Oct 2026* | Apache 2.0 |
| Milvus | Vector DB | 544 → 644 | 56 → 61 | 15 Oct 2026 | Apache 2.0 |
| Weaviate | Vector DB | 832 → 2,533 | 35 → 34 | 15 Oct 2026 | BSD-3 plus Weaviate License (see below) |
| LanceDB | Vector DB | 150 → 468 | 29 → 58 | 15 Oct 2026 | Apache 2.0 |
| pgvector | Postgres extension | 32 → 118 | 3 → 4 | 1 Oct 2026 | PostgreSQL License |
* Qdrant’s default branch, master, is updated at releases; day-to-day work happens on a dev branch, whose latest commit was on 15 October 2026. Haystack’s 2026 figure includes 339 commits from its release bot and dependabot; counting only human-authored commits, it went from 121 to 405.
Commits, a year apart
Non-merge commits on each default branch, 17 Jul to 15 Oct, 2025 vs 2026 (log scale)
Source: EvalAITools analysis of public GitHub history, cloned 15 October 2026.
Three Projects That Changed Course
LlamaIndex: the company’s focus has moved
Commits fell 71% year on year and the number of authors fell 60%. The README explains why, in the company’s own words: its “primary focus has shifted towards LlamaParse”, its document-parsing platform, and the open-source framework is still available “as an open toolkit that you’re welcome to use”. Recent commits are mostly upkeep, such as adding support for new models like Claude Haiku 5.5 and GPT-6.1 Sol. The repository also has a stale-package policy: integrations that score poorly on tests, downloads and commit activity are moved to a separate stale_packages branch, though they stay on PyPI.
What I’d do: existing LlamaIndex apps don’t need to panic. For a new project, check that every integration you plan to use is still on the main branch, and expect slower feature work.
Microsoft GraphRAG: maintenance mode
The README now says the project “is largely in maintenance mode, and won’t be accepting new PRs or implementing new features”, with bug fixes and dependency updates, especially for security issues, as needed. Activity matches: 23 commits from 3 authors in the 2026 window.
What I’d do: GraphRAG is still useful for studying graph-based retrieval, but I wouldn’t start a new production system on it. LightRAG, a graph RAG project that nearly doubled its commits, is the obvious open-source alternative to evaluate.
Weaviate: a new enterprise-licensed folder
On 26 August 2026 Weaviate added a wl directory to its repository under a separate licence. The LICENSE file now says code outside that directory is BSD-3-Clause, while code inside it is “only available under a separate enterprise license”, with the features “enabled via license key”. When we checked, the folder held 21 files covering backup deduplication, database-user expiry, namespaces and a self-recovery system. The core database is still open source, but some newer operational features are not.
What I’d do: if you self-host Weaviate, check whether any feature you need lives under wl/ before you upgrade or plan around it.
Projects That Grew
- RAGFlow quadrupled its commits and went from 93 to 244 authors, the busiest project in this group.
- Haystack more than tripled its human-authored commits and went from 24 to 131 authors.
- txtai went from a single author to 61, as outside contributors joined its long-time maintainer.
- Qdrant and LanceDB both more than tripled their commits and roughly doubled their authors.
- pgvector nearly quadrupled its commits, but it’s still a very small team (4 authors), which is worth knowing if your whole stack sits on it.
Chroma is the one I’d watch: still active, with a commit on 13 October, but its commits fell 62% and its authors halved.
How Much Is AI-Assisted?
We also counted 2026 commits that credit an AI coding tool, usually through a “Co-authored-by” line, or are authored directly by one.
| Project | AI-credited commits (2026 window) | Share |
|---|---|---|
| LightRAG | 875 of 1,476 | 59% |
| DSPy | 96 of 185 | 52% |
| Qdrant | 387 of 970 | 40% |
| Milvus | 247 of 644 | 38% |
| GraphRAG | 6 of 23 | 26% |
| Chroma | 26 of 120 | 22% |
| Weaviate | 487 of 2,533 | 19% |
| LanceDB | 72 of 468 | 15% |
| Haystack | 112 of 744 | 15% |
| RAGFlow | 47 of 2,709 | 2% |
| LangChain | 2 of 552 | under 1% |
| LlamaIndex | 1 of 105 | 1% |
| txtai, pgvector | 0 | 0% |
Claude was the most commonly credited tool, followed by Cursor (mostly in Qdrant), Amp (mostly in DSPy) and OpenAI’s Codex (mostly in Milvus). Disclosure: Anthropic makes Claude, which also helped with this analysis. A low number doesn’t prove a project avoids AI tools; many developers use them without adding a trailer. It mainly tells you about each project’s conventions.
What I’d Choose for a New RAG Project
| If you need… | Consider | Check first |
|---|---|---|
| A general RAG framework | Haystack, LangChain, DSPy | LangChain’s author count fell sharply; read its current docs before relying on older tutorials |
| A full RAG app with a UI | RAGFlow | Its pace of change means frequent upgrades |
| Graph-based retrieval | LightRAG | Avoid starting new work on GraphRAG |
| Vectors inside an existing Postgres | pgvector | Small team; keep an upgrade plan |
| A dedicated vector database | Qdrant, Milvus, LanceDB | Weaviate: whether your features sit in the enterprise-licensed folder |
| An existing LlamaIndex app | Keep it | That your integrations haven’t moved to the stale branch |
Whatever you pick, test retrieval quality on your own documents. My small hybrid search experiment found that the retriever choice mattered less than how documents were cleaned and split.
FAQ
Is LlamaIndex dead?
No. It had 105 commits from 57 authors in the 91 days to 15 October 2026 and still adds support for new models. But commits fell 71% year on year, and the README says the company’s main focus is now LlamaParse.
Is Microsoft GraphRAG still maintained?
Only for bug fixes and dependency or security updates. Its README says it’s largely in maintenance mode and won’t accept new pull requests or features.
Is Weaviate still open source?
Mostly. Code outside the wl directory is BSD-3-Clause; code inside it, added from 26 August 2026, needs a separate enterprise licence and a licence key.
Which vector database is the most active?
By raw commits in our window, Weaviate, then Qdrant and Milvus. By number of contributors, Milvus, LanceDB and Qdrant were close, at 55 to 61 authors each.
Related: Which AI agent frameworks are still maintained · Does hybrid search help RAG? Our test · System 1 vs System 2 AI agents
Sources
- GitHub repositories cloned on 15 October 2026: run-llama/llama_index, langchain-ai/langchain, deepset-ai/haystack, infiniflow/ragflow, HKUDS/LightRAG, microsoft/graphrag, stanfordnlp/dspy, neuml/txtai, chroma-core/chroma, qdrant/qdrant, milvus-io/milvus, weaviate/weaviate, lancedb/lancedb, pgvector/pgvector
- LlamaIndex README and STALE.md; GraphRAG README; Weaviate LICENSE and wl/LICENSE-WEAVIATE, as of 15 October 2026
