Vector databases

What to look for

The questions that separate the options, in order of how much they matter.

1. 01

   Filtered search quality

   Real queries are almost always filtered by permissions, tenant ID, date or document type, and the results must stay relevant once the filter applies. Engines that filter after the vector search can return three results when you asked for ten, or quietly miss the best match. Qdrant adds graph edges for indexed payload fields so filtered traversal stays connected; pgvector needs iterative scans (0.8.0 and later) or partial indexes. Test with your most selective real filter, not an unfiltered demo.

2. 02

   Hybrid search: vectors plus keywords

   Vectors miss names, identifiers and exact terms such as product codes and error messages; keyword search (BM25) catches them. Hybrid search runs both and blends the scores. Weaviate, Qdrant, Milvus, LanceDB and turbopuffer do it in one engine; Pinecone offers sparse-dense indexes or a full-text filter; with pgvector you combine Postgres full-text search in SQL yourself. Check whether you can tune the weighting or use reciprocal rank fusion.

3. 03

   Recall against latency

   Approximate search trades accuracy for speed: recall is the share of the true nearest neighbours it recovers, and you want it at the response time you need, not at any speed. HNSW, a layered graph, gives high recall but wants the index in RAM; IVF clusters vectors and searches the nearest clusters; disk-based indexes (Milvus's DiskANN, Weaviate's HFresh, turbopuffer's and Chroma Cloud's SPFresh or SPANN) keep most data on SSD or object storage. Quantisation shrinks vectors 4 to 32 times, usually with a rescoring step to recover accuracy. On benchmarks: ANN-Benchmarks, run by academics Martin Aumueller, Erik Bernhardsson and Alec Faitfull, plots recall against queries per second for single queries on small public datasets; it compares index algorithms, not databases, and ignores filtering, writes and operations. VectorDBBench is maintained by Zilliz, the company behind Milvus, and adds p99 latency, filtered and streaming cases and cost per query. Compare QPS and p99 at the same recall, and treat any vendor's benchmark of its own product as a claim to reproduce.

4. 04

   Workload capacity

   Size the job before choosing: expected vector count, dimensions, concurrent queries and how often data changes. Vectors times dimensions times bytes gives the raw memory an in-memory index needs. Encore, which sells Postgres hosting, puts pgvector level with Pinecone below about five million vectors and in need of careful tuning beyond that; Chroma suits prototypes to tens of millions; LanceDB's FAQ says the open-source library handles millions of vectors on one node; Milvus and turbopuffer are built for billions.

5. 05

   Updates and deletion

   Check how quickly inserts, edits and deletes become searchable. Some indexes absorb new data immediately, others buffer writes or need a rebuild; turbopuffer writes go to object storage at 248 ms p90 for a 512 KB upsert. Deletion matters for compliance: when a user withdraws a document, it must stop appearing in results promptly, and per-tenant indexes make deleting a whole customer simple.

6. 06

   Security and isolation

   Look for access control, encryption at rest and in transit, and tenant isolation that holds up. If permissions are enforced by a metadata filter, that filter must be reliable on every query, which is another reason to test filtered recall. turbopuffer offers unlimited namespaces and customer-managed keys on enterprise; Qdrant can mark a field as the tenant key; Weaviate has native multi-tenancy, though its namespaces feature now needs a paid licence; in pgvector you partition by tenant or use row-level security.

7. 07

   Operations

   Backups, recovery, monitoring and scaling often cost more than the licence. pgvector adds nothing new to run if you already have Postgres. Qdrant ships as one binary; Milvus at scale needs Kubernetes and a platform team. Chroma and LanceDB run inside your process. Pinecone and turbopuffer are managed only, with bring-your-own-cloud for enterprise buyers. Licence decides your exit: Apache 2.0 (Qdrant, Milvus, Chroma, LanceDB) and the PostgreSQL licence (pgvector) let you self-host freely, Weaviate is open core since September 2026, and the proprietary services mean exporting and re-indexing to leave.

8. 08

   Total cost

   Count indexes, metadata storage, queries and updates, and whether you need a separate database at all. Managed services bill in different units: Pinecone by read units, write units and stored GB, Weaviate Cloud by stored vector dimensions, turbopuffer by usage above a monthly minimum, Chroma Cloud by data written, stored and queried. A read-heavy chatbot and a write-heavy indexing pipeline can rank the same vendors in opposite orders, so model your own volumes on each calculator and against an instance you would run yourself.

Which approach for which data

Different content needs different models. Many systems use more than one and route by data type.

8 tools, compared

Ranked by our rating. Open any review for sources, prices and the full verdict.

Side by side

The specifications that matter for this step. Checked 2026-10-23; confirm on the vendor’s site before you commit.

Connect it to your agent

MCP servers that let assistants like Claude, ChatGPT and Cursor use these tools directly.

• Qdrant MCP Server 3.7
• Supabase MCP Server 4.6
• Neon MCP Server 4.4

Where this stage fits

• Chat with your documents: we would start with Qdrant.
• Customer support agent: we would start with pgvector.
• Coding agent: we would start with LanceDB.
• Research & knowledge graph: we would start with Weaviate.
• Assistant with long-term memory: we would start with Qdrant.
• Image generation pipeline: we would start with Qdrant.
• Video production: we would start with LanceDB.
• Voice agent: we would start with pgvector.
• Private, on-device AI: we would start with LanceDB, Chroma.

Sources for this guide

• ANN-Benchmarks
• VectorDBBench (Zilliz) on GitHub
• VectorDBBench leaderboard
• Qdrant docs: indexing
• Qdrant docs: quantization
• Weaviate docs: vector indexes
• Milvus docs: index types
• Milvus docs: DiskANN
• Chroma docs: collection configuration
• Chroma docs: architecture
• LanceDB docs: indexing
• Pinecone docs: hybrid search
• pgvector README
• turbopuffer: ANN v3
• LanceDB OSS FAQ
• pgvector vs Pinecone (Encore)