Vespa for custom ranking and retrieval

How it works

Vespa applications declare document schemas and ranking profiles. Retrieval can choose candidates before subsequent ranking stages apply more expensive signals.

What you need

Write the application schema and rank profile with a small corpus. Validate document ingestion and query input types before adding ranking functions that depend on those fields.

What to try first

Create an application with text and vector fields, then run the same labeled queries with two ranking profiles. Keep a record of which stage dropped relevant candidates.

Compare a fixed rank profile on queries with common and rare terms. A feature that helps one class may bury the useful result in the other.

Where to look for failures

A more complex ranker only helps if retrieval supplies useful candidates. Separate recall at the candidate stage from final ranking metrics.

Related Franklin pages

Start with a labeled query set and save each engine's candidates. Compare filtered search and ranking behavior before extrapolating throughput from a query without metadata constraints.

• Search embeddings alongside structured metadata in a dedicated database. Qdrant for filtered vector retrieval
• Organize large embedding collections with configurable indexes and search modes. Milvus for managing vector search at scale

For the role of the index in nearest-neighbor search, see Franklin’s glossary. Vector indexes

Project documentation

• Current project README
• README at the reviewed revision