Vector DBs comparaison — pgvector, Qdrant, Milvus, Weaviate

[khalilbenaz]

---

lede: Vector DBs pour RAG : pgvector pour démarrer, Qdrant pour scale OSS, Weaviate pour hybrid search, Milvus pour très massif.
lede_en: Vector DBs for RAG: pgvector to start, Qdrant for OSS scale, Weaviate for hybrid search, Milvus for very massive.
title_en: Vector DBs comparison — pgvector, Qdrant, Milvus, Weaviate

Le marché vector DB

• pgvector : extension Postgres, simple.
• Qdrant : OSS, Rust, performant.
• Weaviate : OSS, hybrid search natif.
• Milvus : OSS, scale massif.
• Pinecone : managed, cher.
• Chroma : dev/prototypes.

pgvector — démarrer

CREATE EXTENSION vector;

CREATE TABLE docs (
  id serial primary key,
  embedding vector(1536),
  content text
);

CREATE INDEX ON docs USING hnsw (embedding vector_cosine_ops);

SELECT content FROM docs
ORDER BY embedding <=> '[0.1, 0.2, ...]'::vector
LIMIT 10;

Sweet spot : moins de 10M vecteurs, déjà Postgres en stack. Pas de service séparé à opérer.

Qdrant — quand pgvector ne suffit plus

OSS, Rust. Performances excellentes. Cloud ou self-host.

from qdrant_client import QdrantClient
client = QdrantClient("localhost", port=6333)
client.search(collection_name="docs", query_vector=embedding, limit=10)

Sweet spot : 10M-1B vecteurs, vous voulez OSS sans dépendre d'un cloud.

Weaviate — hybrid search

Combine vector + BM25 + filters dans une seule query.

client.query.get("Doc", ["content"]).with_hybrid(
    query="red shoes",
    vector=embedding,
    alpha=0.5  # 0 = pure BM25, 1 = pure vector
).do()

Sweet spot : RAG qui veut combine sémantique + lexical (souvent meilleur que vector seul).

Milvus — scale massif

Pour >1B vecteurs. Architecture distributed sérieuse.

Sweet spot : Spotify-scale, vous gérez vraiment des volumes massifs.

Pinecone — managed cher

Service managed. Pas de self-host.

Pricing : à partir de $70/mois pour 10M vecteurs, escalade vite.

Sweet spot : vous voulez zéro ops, budget cloud.

Chroma — dev / prototypes

Python embedded vector store. Rapide à démarrer.

Sweet spot : prototypes, demos, MVPs. Pas pour prod sérieuse.

Decision tree

Déjà Postgres + <10M vecteurs ?
  → pgvector.

OSS + 10M-1B vecteurs ?
  → Qdrant.

Hybrid search important ?
  → Weaviate.

>1B vecteurs ?
  → Milvus.

Cloud managed, budget ok ?
  → Pinecone.

Prototype Python ?
  → Chroma.

Verdict

Pour 90 % des RAG en 2026 : pgvector suffit. Pas de raison de complexifier.

Migrer vers Qdrant/Weaviate quand mesurer limite. Pinecone si budget illimité et zéro ops.

The vector DB market

• pgvector: Postgres extension, simple.
• Qdrant: OSS, Rust, performant.
• Weaviate: OSS, native hybrid search.
• Milvus: OSS, massive scale.
• Pinecone: managed, expensive.
• Chroma: dev/prototypes.

pgvector — start

CREATE EXTENSION vector;

CREATE TABLE docs (
  id serial primary key,
  embedding vector(1536),
  content text
);

CREATE INDEX ON docs USING hnsw (embedding vector_cosine_ops);

SELECT content FROM docs
ORDER BY embedding <=> '[0.1, 0.2, ...]'::vector
LIMIT 10;

Sweet spot: less than 10M vectors, Postgres already in stack. No separate service to operate.

Qdrant — when pgvector isn't enough

OSS, Rust. Excellent performance. Cloud or self-host.

from qdrant_client import QdrantClient
client = QdrantClient("localhost", port=6333)
client.search(collection_name="docs", query_vector=embedding, limit=10)

Sweet spot: 10M-1B vectors, you want OSS without cloud dependency.

Weaviate — hybrid search

Combines vector + BM25 + filters in a single query.

client.query.get("Doc", ["content"]).with_hybrid(
    query="red shoes",
    vector=embedding,
    alpha=0.5  # 0 = pure BM25, 1 = pure vector
).do()

Sweet spot: RAG wanting combined semantic + lexical (often better than vector alone).

Milvus — massive scale

For >1B vectors. Serious distributed architecture.

Sweet spot: Spotify-scale, you really handle massive volumes.

Pinecone — expensive managed

Managed service. No self-host.

Pricing: from $70/month for 10M vectors, escalates fast.

Sweet spot: you want zero ops, cloud budget.

Chroma — dev / prototypes

Python embedded vector store. Fast to start.

Sweet spot: prototypes, demos, MVPs. Not for serious prod.

Decision tree

Already Postgres + <10M vectors?
  → pgvector.

OSS + 10M-1B vectors?
  → Qdrant.

Hybrid search important?
  → Weaviate.

>1B vectors?
  → Milvus.

Cloud managed, budget OK?
  → Pinecone.

Python prototype?
  → Chroma.

Verdict

For 90% of RAG in 2026: pgvector suffices. No reason to complexify.

Migrate to Qdrant/Weaviate when measure limit. Pinecone if unlimited budget and zero ops.

[MikeGarciaAGM]

Implemented and opened PR for this issue: #347

Scope:

• add issue content to posts.json as a new post entry
• preserve existing structure/order