Vector Databases Compared — Qdrant, pgvector, Azure AI Search, Pinecone, Weaviate

Every vector DB looks the same in a demo. They diverge in the cost column, the SDK column, and the "do I want to run this myself" column. After shipping with all five on different projects, here's the unsentimental version.

Store	.NET SDK	Hosting	Hybrid search	1M vec cost/mo	10M vec cost/mo
Qdrant	Official, solid	Self / Cloud	Yes (built-in)	~$25 (1 node)	~$300
pgvector	npgsql + a few lines	Whatever you run Postgres on	DIY	Free if you have Postgres	~$100 (bigger Postgres)
Azure AI Search	First-class	Managed	Yes (built-in RRF)	~$75 (S1)	~$750 (S2)
Pinecone	Decent	Managed only	Sparse-dense yes	~$70	~$700
Weaviate	OK	Self / Cloud	Yes	~$25 self	~$400

The same upsert+search in each (sketched, sparing the imports):

// Qdrant
await client.UpsertAsync("docs", new[] { new PointStruct(Guid.NewGuid(), vec, payload) });
var hits = await client.SearchAsync("docs", qVec, limit: 5);

// pgvector
await db.ExecuteAsync("INSERT INTO docs(id, vec, text) VALUES (@id, @vec, @text)", new {...});
var hits = await db.QueryAsync("SELECT * FROM docs ORDER BY vec <-> @q LIMIT 5", new {...});

// Azure AI Search
await client.UploadDocumentsAsync(new[] { new SearchDocument { ... } });
var hits = await client.SearchAsync("*", new SearchOptions { VectorSearch = new(...) });

The honest recommendation matrix:

Prototype / weekend project: pgvector. You probably already have Postgres. Don't introduce a new service for 50,000 vectors. The "what if we scale" worry is premature. You can migrate later. You probably won't.

Small SaaS (under 5M vectors): Qdrant self-hosted if you already run Docker. Azure AI Search if you're an Azure shop and want it managed. pgvector still works fine up here for many use cases.

Enterprise / Azure shop: Azure AI Search. The hybrid search and RRF are first-class, integration with Azure OpenAI is paved, and your security team already approved it. Cost is meaningful but predictable.

On-prem / data sovereignty: Qdrant. The license is right, the SDK is right, and self-hosting is genuinely smooth.

The one I'd avoid for new projects: Pinecone. It's good. The .NET SDK is decent. But it's managed-only and the cost crosses uncomfortable at 10M+ vectors. Pick this only if you specifically want zero-ops and you're OK with the lock-in.

A note on pgvector that might save you a migration: it's not "the cheap version." For under ~3M vectors with sensible HNSW indexes, it's competitive on latency with the specialised stores. The drop-off comes at scale, with multi-tenant filtering, and when you need real hybrid search (BM25 + vectors with proper RRF). At that point you graduate. Before that, you don't need to.