txtai

NeuML's txtai centers on an embeddings database that pairs dense and sparse vector indexes with a graph layer and SQL-queryable document storage, so one index serves semantic, keyword, graph and SQL-filtered queries. Vectors default to Faiss, with HNSW, Annoy, pgvector, SQLite, NumPy and Torch as alternatives, while content lives in SQLite, DuckDB or an external database reached by an SQLAlchemy URL. Around that core sit pipelines for LLM prompting, extractive QA, summarization, translation, transcription, text to speech and zero-shot labeling, workflows that chain them, and agents built on smolagents. Models come from Hugging Face or local paths, and LLMs can also run through llama.cpp or hosted APIs via LiteLLM. A FastAPI service configured in YAML exposes everything over HTTP and MCP, with bindings for JavaScript, Java, Rust and Go. It needs Python 3.11 or newer and installs with pip or runs in Docker; small embedding models are fine on CPU. Apache 2.0 licensed and usable commercially, with optional consulting and hosting sold by NeuML. Fits Python developers wanting RAG, search and agents in one local-first package.