Source: neuml/txtai on GitHub (Apache-2.0, about 13.4k stars), highlighted in posts by André Lindenberg and Stanislav Beliaev, 2026-11-16.
What it is
txtai is an open-source Python framework from NeuML built around an "embeddings database": one index that unifies sparse and dense vector indexes, a relational store, and an optional graph network. It runs locally by default and adds pipelines (LLM prompts, QA, summarization, transcription), workflows, agents, and Web and MCP APIs on top.
The parts worth a staff engineer's attention
One query for meaning and structure. Its SQL dialect includes a
similar()clause, so a single statement can rank by semantic similarity while filtering on ordinary columns such as dates or tags. That removes the usual two-hop pattern of vector search followed by a metadata filter in another store, and the consistency questions between them.A semantic graph as an optional layer. Indexing can link each document to its nearest neighbours, and an LLM-driven entity-extraction step can add typed relationships. That gives topic exploration and graph-walk retrieval (the repo has GraphRAG examples) without a separate graph database.
Local-first and small. Install with pip, start with a few lines, and add dependencies only when needed. Bindings exist for JavaScript, Java, Rust and Go via the API.
Where it fits
A good default for single-node or embedded retrieval: internal knowledge bases, agent memory stores, notebooks, and prototypes where operating a vector database plus a relational database plus a graph store is too much. It is less obviously the answer for multi-tenant, high-write, horizontally scaled search, where a dedicated engine's sharding and operations tooling matter more.
Caveats
Star counts and feature claims are from the project and the posts; they are not independently benchmarked.
LLM-extracted relationships inherit the extractor's errors, so treat graph edges as hints unless they are validated.
NeuML sells consulting and a hosted offering around the stack.
Related entries
R096 and other retrieval entries under index subtopic 4.3; agent memory under 1.1.