Vespa vs Qdrant vs Turbopuffer for large-scale hybrid search (BM25 + text & image vectors)

[sumitrajrajsumit]

Hi everyone — we’re evaluating search platforms for a hybrid search use case and would appreciate insights from people who’ve used Vespa, Qdrant, or Turbopuffer in real systems.

Background

• Use case: B2B product search
• Catalog size: ~170 million products
• Traffic: ~100 QPS (steady-state, higher during peaks)
• Latency goal: sub-100 ms end-to-end
• Data: each product has text and images

Search requirements:

• Traditional keyword search (BM25 or equivalent)
• Text vector search
• Image vector search
• Ability to combine these signals into a single ranked result set
• Current system: Apache Solr
• Deployment model: open (self-hosted or managed)

What we’re trying to understand

Hybrid search capabilities

• How mature is support for combining BM25 and vector search?
• How flexible is the scoring and ranking pipeline?

Scalability & reliability

• Experiences operating at hundreds of millions of documents
• Any architectural or operational challenges at this scale?

Business logic & reranking

• How easy is it to apply custom business logic (boosting, filtering, reranking)?
• Is this typically handled in-system or via an external reranking service?

Performance & latency

• Realistic latencies for hybrid retrieval at ~100 QPS
• Trade-offs between accuracy, recall, and speed

Feature set & ecosystem

• Query language expressiveness, filtering, multi-vector support
• Tooling for relevance tuning, monitoring, and operations

Cost considerations

• General cost patterns (managed vs self-hosted)
• How costs scale with data size and query volume

Migration perspective

• Lessons learned when migrating from Solr or Elasticsearch

Open question
If you were designing a hybrid search system today for ~170M products, ~100 QPS, and sub-100 ms latency, how would you think about choosing between Vespa, Qdrant, and Turbopuffer? What factors would matter most?

Any references to documentation, benchmarks, or real-world experiences would be greatly appreciated.

[azayarni]

We can help you design such a solution with Qdrant. We have several customers with similar requirements.
Submit your request here and our team will get in touch https://qdrant.tech/contact-us/