Skip to content

Repo Boundaries

This repo is a public sample accelerator for Azure AI Search / Foundry IQ live Knowledge Sources. It is not a production reference architecture and it is not a Fabric ontology authoring guide.

In Scope

  • Synthetic Airline Operations sample data.
  • Offline retrieve responses that demonstrate the trace contract.
  • MCP Server Knowledge Source payloads and walkthroughs.
  • BYO Fabric Ontology Knowledge Source validation.
  • Combined Knowledge Base routing examples.
  • Deployment automation for Azure AI Search, Azure OpenAI, app hosting, and sample Knowledge Source assets.
  • Public-preview caveats and validation guidance.

Out Of Scope

  • Internal Fabric ontology authoring steps.
  • Private preview setup, allowlisting, or unpublished tenant instructions.
  • Customer data, real tenant IDs, real workspace IDs, real ontology IDs, API keys, bearer tokens, raw live responses, or screenshots with sensitive values.
  • Recommending Fabric MCP through MCP Server Knowledge Source as the primary path. Use native Fabric Ontology Knowledge Source for the Fabric path.
  • Production hardening claims that are not backed by explicit evidence.

Agent Rules

  • Start with offline replay unless the user asks for live resources.
  • Prefer mcp-only before Fabric paths.
  • Use byo-fabric only when existing Fabric workspace and ontology IDs are available.
  • Treat full as maintainer/demo-oriented because it can create Fabric capacity and sample Fabric assets.
  • Keep generated reports and logs in ignored paths.
  • Summarize live evidence with sanitized status and counts, not raw tenant payloads.

Evidence Standards

Use the right evidence for the claim:

  • Repo shape: bash scripts/validate-local.sh
  • Offline trace shape: samples/responses/*.sample.json
  • Route expectations: evals/expected_routes.yaml
  • Live MCP path: mcp-only deployment or E2E evidence
  • Live Fabric path: byo-fabric deployment or E2E evidence with delegated source authorization
  • Greenfield demo path: full mode evidence plus cleanup evidence

Offline replay is useful learning evidence. It is not proof of live Fabric retrieval.