Qdrant vs Weaviate vs Milvus vs Pinecone: what to compare

Start here

Your search demo finds a useful paragraph. The harder questions arrive when you add real filters, more documents and a recovery requirement. Before selecting a vector database, write down those conditions. This guide helps you design a comparison around retrieval quality and operational work, without pretending one benchmark fits every application.

Begin with the retrieval problem

The linked documentation describes the platforms and their indexing concepts. Before comparing implementation choices, define your records, vector dimensions, metadata filters and expected query mix. A demonstration with unfiltered vectors does not establish performance for filtered production traffic.

Make the measurements comparable

Use the same dataset and relevance evaluation across candidates. Record configuration changes that trade retrieval quality for speed. Keep a reproducible query set so a faster run is not simply an easier one.

• Data volume, vector dimensions and metadata distribution.
• Target retrieval quality and number of results returned.
• Concurrent requests, update rate and filtering conditions.
• Index build time, steady-state memory and recovery behavior.

Count operating costs beyond the index

For self-hosting, include the infrastructure and the work of upgrades, backups and incident response. For a managed service, calculate the current billed units for your workload, including storage and reads or writes where applicable. Compare a month of expected usage; this page intentionally does not reuse unverified dollar or p99 latency figures.

A sensible shortlist process

Start with the platform your team can operate competently, then test the requirement most likely to disqualify it. A useful proof of concept includes a restore or rebuild, not only a successful query. Keep the configuration and evaluation sheet with the decision so the choice can be revisited as volume changes.

Try this next

Check the other half of retrieval quality: whether your extraction pipeline preserves the source material you need. Read Firecrawl or Playwright for RAG? Choose the missing step.