Skip to main content
A.I. Commerce Studio

Resources · Integration patterns

RAG ingestion and retrieval

Documents are indexed with their permissions and the model answers only from permitted, cited passages.

Integration status: Documented pattern

A documented pattern. The product has no RAG index.

Description

Documents are chunked, embedded and stored with metadata and permissions. On a question, retrieval filters by permission and the model answers only from the passages, naming the source.

Prerequisites

  • Curated documents with owners.
  • A permission model.
  • A test set of questions.

When to use it

  • Knowledge in documents that change.
  • Answers need a source.
  • Different access rights.

When not to use it

  • The data is structured — use queries.
  • Documents have no owner.
  • The accuracy needed requires computation.

Reference diagram

An illustration of the components and the flow — not a running system.

  1. 01Ingestion

    • Sources
    • Chunking
    • Metadata and permissions
  2. 02Index

    • Vectors
    • Keywords
  3. 03Answer

    • Permission filter
    • Model
    • Citation
The diagram as a list
StageComponents
IngestionSources, Chunking, Metadata and permissions
IndexVectors, Keywords
AnswerPermission filter, Model, Citation

Security controls

  • Permission filtering before the model.
  • Input treated as untrusted.
  • No question content in logs.

Observability

  • Share of answers with a source.
  • Questions without results.
  • Index freshness.

Failure modes

FailureMitigation
Answer from an outdated documentDates, owners and periodic clean-up.
Leak across permissionsFilter in retrieval, tests with roles.
Made-up answerNo source — no answer.