Making Azure AI Foundry Agents Explainable — Knowledge Graphs + Source Attribution

[BIBIN PRATHAP (bibinprathap)]

Technical Feedback

Describe the precise feedback.
90% of Azure AI demos work on stage — most never ship. The gap is architecture, not the model. I put together a field guide on taking Azure AI Foundry agents from POC to production, using knowledge graphs (GraphRAG) for grounding and source attribution.

What the write-up covers:

• Grounding with a knowledge graph — multi-hop retrieval + inline citations so every answer is traceable to a source (critical for regulated/medical use cases).
• Externalized state — keeping agent memory and session state outside the model.
• Identity at the boundary — Entra ID / RBAC instead of trusting the prompt.
• Observability & compliance — tracing, evaluation, and auditability.
• The stack — Azure AI Foundry + Model Context Protocol (MCP) + Azure Functions (Flex Consumption) + Azure OpenAI, from a real build (VeritasGraph medical MCP server).

📖 Full write-up: https://bibinprathap.com/blog/azure-ai-proof-of-concept-to-production
▶️ 6-min walkthrough: https://youtu.be/z-CPS5WUvyw?si=fmpo1RN28Bh9KbIb

Desired Outcome

Describe the desired outcome.
I'd like to compare notes with the community and see clearer first-party guidance/patterns from Foundry for production-grade grounding and auditability. Specifically:

1. How are you grounding Foundry agents today — vector RAG, GraphRAG, or hybrid?
2. Anyone combining knowledge graphs with MCP tools in Foundry? What worked / broke?
3. For regulated domains, how are you handling source attribution and audit trails?

Requested solution: reference architectures / samples in Foundry for GraphRAG + MCP grounding, plus built-in support for inline source attribution and audit trails on agent responses.

Current Workaround

Describe the current workaround.
I built a custom MCP server (VeritasGraph, a medical MCP server) on Azure Functions (Flex Consumption) that exposes GraphRAG tools to a Foundry agent:

• Knowledge-graph grounding with multi-hop retrieval and [doc#chunk] citations for every answer.
• Agent memory/session state kept outside the model.
• Entra ID / RBAC enforced at the boundary.
• Tracing + evaluation for observability and auditability.

Stack: Azure AI Foundry + MCP + Azure Functions + Azure OpenAI.
Happy to share more detail on the GraphRAG retrieval design if useful.
source code:
https://github.com/bibinprathap/azure-ai-foundry-veritas-graph