The outcome

Return a small, well-ordered evidence set that measurably serves real queries across required languages.

Step by step

A workflow you can repeat.

  1. 01

    Define users, languages, corpus, source authority, chunking, freshness, access rules, query classes, relevance labels, latency, recall, precision, and cost targets.

  2. 02

    Choose a compatible embedding and reranker, record model versions, dimensions, context, task settings, license, and build a hard evaluation set with judged positives and negatives.

  3. 03

    Normalize and chunk authorized documents without losing titles, URLs, dates, permissions, or section boundaries, then embed and index versioned content.

  4. 04

    Retrieve a broad candidate set and rerank a smaller set, measuring each stage separately for language balance, false positives, missing evidence, latency, and cost.

  5. 05

    Pass only cited excerpts to generation, treat retrieved text as untrusted data, enforce access at query time, and reevaluate before any model, chunking, or corpus change.

Working standard

What good use looks like.

  • Measure retrieval and reranking separately.
  • Preserve source and permission metadata.
  • Re-embed after model changes.

Official references

Check the current product documentation.