Bases de données vectorielles auto-hébergées

Réponse rapide

La base de données vectorielle est le milieu d'une pile RAG auto-hébergée : elle stocke les vecteurs qu'un modèle d'embedding produit et, au moment de la requête, trouve les correspondances les plus proches pour qu'un LLM local y réponde. Ceci compare 14 bases de données vectorielles auto-hébergeables selon l'infrastructure requise (intégrée à votre application, une extension Postgres, un serveur unique, ou un cluster), leurs types d'index, si elles font de la recherche hybride NATIVE (dense + mots-clés), la quantification, et — la colonne qui compte le plus pour la souveraineté — si la licence est véritablement open source ou seulement « source-available ». Trié par friction de déploiement, du plus simple au plus complexe.

Choix honnêtes pour 2026 : les démarrages à moindre friction sont sqlite-vec (une extension SQLite), Chroma (pip install), ou LanceDB (intégré ET hybride BM25+vecteur natif) ; si vous exécutez déjà Postgres, pgvector garde vecteurs, métadonnées et jointures dans une seule base ACID. Pour un serveur de production à binaire unique avec recherche hybride native, Qdrant (Rust, Apache-2.0) ou Weaviate (BSD-3) sont les choix proprement ouverts ; Milvus et Vespa passent à l'échelle des clusters. Attention licences — la colonne que ces listes ratent d'habitude : Elasticsearch N'EST PAS open source (chaque build téléchargeable est sous licence Elastic / SSPL ; son AGPL ne couvre que la source) — utilisez OpenSearch (Apache-2.0) comme équivalent réellement ouvert. Redis a rouvert son produit livré sous AGPLv3 en v8 (le copyleft réseau s'applique), et VectorChord est en AGPL/ELv2 (une régression d'ouverture par rapport à son prédécesseur pgvecto.rs sous Apache-2.0). Et « hybride » ici signifie fusion dense+creux native — pgvector, sqlite-vec et FAISS vous obligent à la fusionner vous-même. CSV/JSON gratuits sous CC BY 4.0.

Télécharger le CSV Télécharger le JSON API REST →

OSI ouvert = véritable licence open source ; source-available = usage restreint (SSPL/BSL/Elastic/RSAL). Hybride = fusion dense+creux native, pas du filtrage de métadonnées. ~ = confiance moindre.

Base de données	Type	Langage	Index	Hybride	Quant.	Licence	Idéal pour
Chroma (ChromaDB) Chroma (chroma-core)	embedded-library	Rust core (1.x rewrite) with Python/JS clients	HNSW (SPANN in the distributed/Cloud engine)	partial	none (stores full-precision vectors)	Apache-2.0 Core is Apache-2.0 (cleanly OSI-open). Hybrid caveat, not license: stable self-hosted OSS does dense + metadata + document $contains/$regex full-text FILTER, not BM25 fusion; native sparse (BM25/SPLADE) + dense RRF 'Search() API' is newly announced and documented Cloud-first — OSS-GA unconfirmed.	fastest 'pip install and go' DX for a RAG prototype; embedded first, scales to a single server later
LanceDB LanceDB Inc. (built on the Lance columnar format)	embedded-library	Rust (with Python, TypeScript, Rust, Go SDKs)	IVF-PQ, IVF-HNSW variants (HNSW), flat/brute-force; disk-based columnar Lance format	yes	product (PQ), scalar (SQ), binary	Apache-2.0 Embedded OSS core is Apache-2.0 and includes native FTS+vector hybrid; LanceDB Cloud/Enterprise are separate proprietary offerings (does not affect the OSS boolean).	embedded multimodal RAG at scale on a laptop or object storage — native Tantivy BM25 + vector hybrid with rerankers, no server to run
sqlite-vec Alex Garcia (asg017); backed by Mozilla Builders	embedded-library	C (zero dependencies)	flat/brute-force KNN in stable v0.1.9 (2026-03-31); IVF ('experimental, not enabled') + DiskANN only in v0.1.10-alpha pre-releases; no HNSW. vec0 virtual tables with partition & metadata columns	no	binary (bit vectors) + scalar (int8); float32 default	MIT OR Apache-2.0 (dual)	lowest-friction option — add vector search to an existing SQLite/libSQL app with one extension load, no server; brute-force only at GA so best for small/medium corpora
txtai NeuML (David Mezzetti)	embedded-library	Python	pluggable ANN backends: Faiss (default), HNSW (hnswlib), Annoy, NumPy brute-force, Torch, pgvector, sqlite-vec	yes	scalar; product/scalar via the Faiss backend	Apache-2.0	all-in-one embedded RAG/semantic-search framework — native sparse(BM25)+dense hybrid, SQLite content store, pick your ANN backend
pgvector Andrew Kane	database-extension	C	HNSW, IVFFlat, flat/exact	no	scalar (halfvec/fp16), binary (bit), plus sparsevec type	PostgreSQL License PostgreSQL License is an OSI-approved permissive (MIT/BSD-style) license — cleanly open.	self-hosters who already run Postgres — keep vectors, metadata, and joins in one ACID database (native BM25 fusion is DIY via tsvector + manual RRF)
VectorChord (vchord) TensorChord (successor to pgvecto.rs)	database-extension	Rust (pgrx)	RaBitQ-based IVF, disk-friendly ANN (DiskANN-style); pairs with the vchord_bm25 companion extension for keyword ranking	partial	binary (RaBitQ), scalar	AGPL-3.0-only OR Elastic-License-2.0 (dual) VERIFIED dual AGPL-3.0 OR ELv2 (both VectorChord and the vchord_bm25 companion). license_osi_open=true only because AGPLv3 is genuinely selectable and copyleft-open — NOT permissive, and a real openness downgrade from its pgvecto.rs predecessor (Apache-2.0). The ELv2 option is source-available/non-OSI. Hybrid is 'partial': requires the separate vchord_bm25 extension + manual fusion.	Postgres users who need pgvector-compatible search at larger scale / lower cost via RaBitQ disk indexing — but note AGPL/ELv2 dual licensing, not the permissive Apache-2.0 of pgvecto.rs
Qdrant Qdrant Solutions GmbH	single-binary-server	Rust	HNSW (payload-aware / filterable); flat/brute-force for exact search	yes	multiple (scalar, product, binary)	Apache-2.0 Core DB (incl. sparse+dense hybrid via the Query API) is Apache-2.0. Some cloud/enterprise features (e.g., RBAC) are proprietary but do not gate the OSS vector+hybrid engine.	purpose-built vector search with rich payload pre-filtering, native sparse(BM25/BM42)+dense fusion (RRF/DBSF), and an easy single-binary self-host
Redis (Query Engine / Vector Sets) Redis Ltd. (originally Salvatore Sanfilippo / antirez)	single-binary-server	C	HNSW and FLAT (brute-force) — used by both the Query Engine vector index and the Vector Sets type	partial	scalar int8 (Q8) and binary in the Vector Sets type; classic Query Engine FLOAT32/FLOAT16 vectors are unquantized	Tri-license (Redis 8+): AGPLv3 OR RSALv2 OR SSPLv1 VERIFIED. Timeline: BSD-3 → (Mar 2024) dual RSALv2/SSPLv1 source-available, non-OSI → (May 2025, Redis 8.0) AGPLv3 re-added as a third option. license_osi_open=true because — unlike Elasticsearch — AGPLv3 covers the ACTUAL Redis 8 product you build/run, and vector search (Query Engine, formerly the RediSearch module, + the new Vector Sets type) is folded into Redis 8 core under the tri-license. RSALv2 and SSPLv1 remain non-OSI options; pre-8, vector search was a source-available module only. AGPL network-copyleft still applies.	ultra-low-latency in-memory vector search co-located with caching for real-time RAG
Weaviate Weaviate B.V.	single-binary-server	Go	HNSW, flat/brute-force, dynamic (flat auto-upgrades to HNSW)	yes	multiple (product PQ, binary BQ, scalar SQ, rotational RQ)	BSD-3-Clause Core is BSD-3-Clause (permissive, cleanly OSI-open) — genuinely open, no source-available trap on the DB core. Native BM25F+dense fusion (rankedFusion/relativeScoreFusion) built in.	developer-friendly native BM25F + dense hybrid search with built-in vectorizer modules
OpenSearch (k-NN plugin) OpenSearch Software Foundation (Linux Foundation); forked from Elasticsearch by AWS	search-engine	Java	HNSW (Lucene / faiss / nmslib engines), IVF (faiss), flat/exact k-NN	yes	multiple (byte/scalar, faiss product PQ, binary/BQ, FP16 disk-based)	Apache-2.0 Apache-2.0 and now under vendor-neutral Linux Foundation governance (OpenSearch Software Foundation, 2024). THE truly-open, permissive alternative to source-available Elasticsearch — the key sovereignty pick in the ES/OpenSearch split.	Apache-2.0 full search engine + native BM25 + kNN hybrid (normalization processor); the permissive Elasticsearch alternative
Elasticsearch Elastic N.V.	search-engine	Java	HNSW (Lucene), flat/brute-force exact kNN, int8/int4/BBQ-quantized HNSW	yes	multiple (int8 scalar, int4, BBQ better-binary-quantization)	Elastic License 2.0 / SSPL 1.0 (all downloadable binaries) + AGPLv3 source-only option (added 8.16, 2024) CORRECTED to FALSE (candidate had true). Every official downloadable distribution is Elastic License 2.0 / SSPL 1.0 — both source-available, NON-OSI. Elastic added an OSI-approved AGPLv3 option in 8.16 (announced 2024) but per Elastic's own licensing FAQ, 'Elastic does not currently offer any downloadable distribution that is covered by the AGPL' — AGPL applies to source code only, so you must self-compile to get an OSI-open build. Practical/mainstream posture remains source-available; contrast Redis 8 (AGPL covers the shipped product) and OpenSearch (clean Apache-2.0).	mature search engine with native RRF hybrid (BM25 + dense kNN + ELSER sparse) at enterprise scale — but source-available licensing; choose OpenSearch if you need a permissive OSI license
Milvus (incl. Milvus Lite embedded mode) Zilliz / LF AI & Data Foundation	distributed-cluster	Go, C++ (Milvus Lite: Python wrapper over the C++ core)	Full Milvus: HNSW, IVF (FLAT/SQ8/PQ), DiskANN, SCANN, FLAT, GPU (CAGRA/GPU_IVF), SPARSE_INVERTED_INDEX. Milvus Lite embedded: FLAT + IVF_FLAT (dense) and SPARSE_INVERTED_INDEX (sparse) ONLY — no HNSW/DiskANN locally	yes	multiple (scalar SQ8, product PQ, binary) in full Milvus; limited in Lite	Apache-2.0 Apache-2.0 across full Milvus and Milvus Lite (same project/repo — the two research-pass rows were deduped into this one entry). Native dense+sparse+BM25 hybrid with RRF is confirmed even in Lite.	billion-scale distributed vector search with the widest index selection and GPU indexing; Milvus Lite gives the same API for embedded local prototyping that lifts to Standalone/Distributed/Zilliz Cloud unchanged
Vespa Vespa.ai (spun out of Yahoo!)	distributed-cluster	Java, C++	HNSW (real-time modifiable), multi-vector HNSW, flat/brute-force	yes	multiple (int8, binary/hamming via tensor cell types)	Apache-2.0	very large-scale hybrid retrieval (text + tensor in one ranking expression) with complex multi-phase ML ranking in a single engine
FAISS (Facebook AI Similarity Search) Meta (Fundamental AI Research)	library-only	C++ (with Python/NumPy bindings; some GPU-accelerated)	flat, IVF, IVF-PQ, HNSW, LSH, IVF-SQ, PQ/OPQ, IVF-HNSW composites	no	product (PQ/OPQ), scalar (SQ), binary	MIT Not a database — a raw ANN library. db_type intentionally 'library-only'; no persistence/metadata/serving layer of its own.	the raw ANN engine underneath many of the above — maximum index control when you'll build the storage/metadata/serving layer yourself

C'est la couche de stockage d'une pile RAG auto-hébergée. Alimentez-la avec des vecteurs d'un modèle d'embedding local (étape 1) et répondez avec un runtime de la comparaison des runtimes IA (étape 3). Fait partie de la base de connaissances IA locale de D-Central. Les licences et fonctionnalités évoluent vite — confirmez la LICENSE actuelle de chaque projet avant de vous engager.

Produits, réparations et guides connexes

Dernière révision: 27 juillet 2026.