Qdrant vs. pgvector: which vector DB should you pick for RAG?

Building a RAG system you hit an early question: dedicated vector DB like Qdrant, or just PostgreSQL with pgvector? Both work — but they have very different strengths.

Quick comparison

Aspect	Qdrant	pgvector
Setup effort	Separate container	Postgres extension
Performance above 1M vectors	Excellent	OK with IVFFlat/HNSW
Metadata filtering	Very fast (payload index)	Flexible via SQL, slower
Hybrid search (vector + BM25)	Built-in	Manual via tsvector
Pure vector operations	Optimised	Bolted onto Postgres
Backups / replication	Own mechanisms	Standard Postgres tooling
Ops cost	A second DB	Just Postgres

When Qdrant?

• You need fast retrieval at 5M plus vectors
• You use complex payload filters (e.g. lang=en AND date>2025)
• You want hybrid search out of the box
• You build an AI-native product where vector search is core

When pgvector?

• You already run Postgres and want to avoid a second DB
• Dataset is under 1M vectors
• Your ops team knows Postgres but not Qdrant
• You need transactional consistency between vectors and metadata

A performance note from real projects

In a project with ~750k legal document chunks, both DBs delivered comparable latencies under 50ms in our setup. The real bottleneck was not the DB, it was:

1. Embedding model (small vs. large model)
2. Chunk size (200 vs. 800 tokens)
3. Hybrid strategy (RRF with BM25 gave +15% Recall@5)

Hybrid search with RRF

In production we combine vector search with BM25 via Reciprocal Rank Fusion:

def rrf(results_lists, k=60):
    scores = {}
    for results in results_lists:
        for rank, doc_id in enumerate(results):
            scores[doc_id] = scores.get(doc_id, 0) + 1 / (k + rank)
    return sorted(scores.items(), key=lambda x: -x[1])

In Qdrant this is built in (prefetch + fusion). With pgvector you have to run tsvector search and vector search separately and fuse the results yourself.

Recommendation

• Prototype & under 500k vectors: pgvector – simpler, one less DB
• Production & over 1M vectors + complex filters: Qdrant
• Hybrid search is critical: Qdrant

Both are solid. The DB is rarely the bottleneck — embedding model, chunking and retrieval strategy matter more.