// the find
nuclia/nucliadb
NucliaDB, The AI Search database for RAG
NucliaDB is a search database built for RAG pipelines that combines vector, full-text, and graph indexes in a single system. The core index engine (nidx) is written in Rust with Tantivy under the hood; Python handles the API layer and orchestration. It's aimed at teams building document search on unstructured data who don't want to stitch together separate vector and keyword search systems.
The nidx Rust layer is genuinely well-structured — separate crates per index type (nidx_paragraph, nidx_text, nidx_relation, nidx_vector) means each concern compiles and tests independently. Running real hybrid search — vector, BM25, and graph traversal in one query — without gluing three separate systems together is the actual value here. The Python binding exposes the Rust core cleanly, so you get Rust performance on hot paths without abandoning Python's ML ecosystem. CI includes e2e tests, codecov, and OSV scanning, which is more pipeline discipline than most projects at this star count.
AGPL v3 is the first thing to read before adopting: if you run a modified version as a service, your modifications must be open-sourced. The README's gloss on this ('free to use as long as you don't modify') understates the constraint for anyone building a product on top. The self-hosted deployment story is operationally heavy — seven separate Helm charts (writer, reader, ingest, search, train, nidx indexer, nidx scheduler) means you're running a small platform, not dropping in a database. The most interesting features (Understanding API for extraction, Learning API for model training) are cloud-only and hit Nuclia's commercial service; the open-source version is meaningfully narrower than the marketing copy implies. Community is thin at 719 stars versus Qdrant, Weaviate, or Milvus — which matters because troubleshooting beyond the docs means reading source or asking in a small Slack.