Open Data Infrastructure

AI-Ready Data Contracts for Vector Indexes

The data contracts vector indexes need for lineage, freshness, policy status, provenance, and evaluation evidence.

A vector index is not a truth layer. It is a derived data product with amnesia unless you build the memory back in.

The practical problem

Vector indexes often sit outside the governed data platform. Source documents are chunked, embedded, filtered, reranked, and served to models. Somewhere in that process, teams lose source identity, freshness, policy status, or provenance.

That is not an AI problem. It is a data contract problem. The index needs a contract just like any other data product.

The contract starts before embedding

An AI-ready data contract for a vector index should name the source, owner, allowed purpose, sensitivity label, refresh rule, chunking method, embedding model, index version, and deletion behavior. The contract should also define what happens when source access changes.

The embedding is derived data. The chunk is derived data. The nearest-neighbor result is derived data. Each step needs enough metadata to preserve meaning and governance context.

Core idea: a vector index should inherit data product controls instead of creating a parallel, ungoverned retrieval path.

Evaluation evidence belongs in the contract

OpenAI evaluation guidance emphasizes structured tests for AI system behavior. For retrieval systems, the evaluation set should include expected sources, disallowed sources, freshness cases, policy denials, and known failure examples.

That evidence should connect to the index version. If an answer changes after re-embedding, the team should know whether the change came from source data, chunking, embedding, ranking, policy filtering, or model behavior. That is retrieval governance in practice.

What breaks first

  • Chunks lose the source owner and policy classification.
  • Deleted or restricted documents remain embedded in the index.
  • Evaluation results track answer quality but not retrieved source quality.
  • The index version changes without a consumer migration plan.

Questions to ask

  • What source contract does each chunk inherit?
  • Which policy checks happen before and after retrieval?
  • How are index versions tied to evaluation results?
  • Can the system remove or quarantine derived context when source access changes?

For adjacent reading, use Foundation Models Need Data Contracts, AI-Ready Data Evaluation Sets, and Data Modeling for RAG and Structured Retrieval.

Sources to start with

These primary sources anchor the technical claims in this guide.

The index can only be trusted when the contract survives the embedding.

ODI hub Article library Use the scorecard Foundation data contracts Retrieval governance AI-ready eval sets