Skip to content

Offline Replay

Offline replay lets you inspect expected retrieve shapes without live Azure resources, Fabric tenant access, or Microsoft Learn MCP calls.

Why It Exists

Live Knowledge Sources are tenant and network dependent:

  • MCP Server KS needs an HTTPS MCP endpoint that Azure AI Search can reach.
  • Fabric Ontology KS needs a Fabric workspace, ontology item, and end-user source authorization for live retrieval.
  • Combined Knowledge Bases can route differently depending on source descriptions, query wording, and model behavior.

The checked-in responses are small synthetic examples for learning the trace structure. They are not captured customer data.

Run The Inspector

python samples/python/inspect_retrieve_response.py samples/responses/mcp-retrieve.sample.json
python samples/python/inspect_retrieve_response.py samples/responses/fabric-airline-ops-retrieve.sample.json
python samples/python/inspect_retrieve_response.py samples/responses/combined-airline-ops-retrieve.sample.json

What To Look For

For MCP Server KS:

  • activity[*].type == "mcpServer"
  • activity[*].toolName or activity[*].mcpServerArguments.toolName
  • MCP references with sourceData

For Fabric Ontology KS:

  • activity[*].type == "fabricOntology"
  • activity[*].fabricOntologyArguments.search
  • Fabric references with sourceData.fabricAnswer
  • Fabric references with sourceData.fabricRawData

For a combined Knowledge Base, treat knowledgeSourceParams as runtime options, not as a strict source allow-list. If you need deterministic validation, run a single-source Knowledge Base first.