What Happened
Vespa 8.786.55 can expose schema guidance to AI agents through /skills/v1/ when the schema skill is enabled in its service configuration. Teams can add explanations to schemas, fields, and rank profiles; Vespa incorporates those explanations into guidance that changes with the schema [1].
In a Vespa evaluation of 196 data-dependent tasks and variants, an agent using the full guidance answered 93% correctly, compared with 9% when explanations were omitted. Multi-turn tests also showed fewer queries, less model time, and fewer tokens. On the STaRK-Amazon product-search dataset, the guidance improved mean reciprocal rank by 12% and reduced token cost by 68%, averaged over three runs [1]. These are workload-specific results, not a guaranteed improvement for every RAG application.
Why It Matters to Businesses
Retrieval quality depends on more than embeddings. An agent must know which fields to filter, which ranking options to use, and what business terms mean. A vector database can find semantically similar items while still returning the wrong answer if, for example, the agent misinterprets whether a price includes VAT. Making those conventions explicit can reduce failed queries and unnecessary model calls [1].
Kimbodo Engineering Perspective
We would treat schema guidance as an interface contract for retrieval, not as a replacement for search evaluation. The useful pattern applies across RAG stacks built with LlamaIndex, LangChain, Haystack, or direct search APIs: document field semantics and permitted query patterns close to the index, then test whether agents actually use them. Vespa’s measured gains support that approach, but do not establish equivalent gains for Weaviate, Pinecone, Qdrant, Milvus, or Elasticsearch [1].
How We Would Implement It
- Define the search contract: searchable fields, filters, ranking profiles, units, null handling, and business definitions such as tax treatment.
- For Vespa, enable the schema skill and add concise explanations where field names or ranking behavior are ambiguous [1]. For other engines, supply equivalent versioned guidance to the application’s retrieval layer.
- Constrain the agent to approved query templates or validated search APIs rather than allowing unrestricted query construction.
- Evaluate representative questions against known answers; measure retrieval relevance, answer correctness, query count, latency, and token cost before and after changes.
- Update guidance and tests whenever schemas or business definitions change.
Risks, Costs and Security
Publishing schema guidance may reveal field names and ranking behavior; restrict access if that information is sensitive. Treat retrieved content as untrusted data, validate agent-generated filters and query parameters, and enforce document-level permissions in the search service—not in prompts alone. Guidance also needs maintenance: stale explanations can make a valid query confidently wrong. The implementation cost is usually smaller than an index migration, but production value depends on evaluation coverage and ongoing schema governance.
Where Kimbodo Comes In
Kimbodo builds and operates this in production for businesses — see our RAG Development Services practice, or Estimate My RAG System.