MENU
A customer support bot built with Azure AI Foundry Agent Service must ground its answers strictly in an internal knowledge base and avoid hallucinating unsupported facts.
What should you implement?
Connect the agent to an Azure AI Search knowledge source and explicitly constrain its instructions to answer from retrieved evidence. Microsoft Foundry Agent Service can use Azure AI Search to retrieve proprietary organizational content and ground model responses in that content. Retrieved documents provide the factual context from which the model generates its answer, and Azure AI Search can return source references for attribution.
Microsoft's RAG guidance also specifically recommends using clear system instructions that require the model to stay within retrieved content when hallucination risk must be reduced. If the available evidence does not support an answer, the instructions should direct the agent to state that the information is unavailable rather than inventing facts. Retrieval grounding substantially reduces unsupported generation, although production systems should still evaluate outputs because grounding alone does not mathematically guarantee zero hallucinations.
Increasing temperature would generally increase output variability and is contrary to the objective. A larger context window without retrieval does not give the model access to the organization's internal knowledge base. Disabling content filtering weakens safety controls and has no role in factual grounding.
Study Guide references: Microsoft Foundry Agent Service Azure AI Search tool; RAG grounding; knowledge sources; source attribution; grounding instructions.
===============