Skip to content

Azure Inference Endpoint Probe Audit

  • Audit date: 2026-03-18 (Asia/Tbilisi)
  • Repository root: /home/standard/dspy_rag_in_repo_docs_and_impl1
  • Git HEAD during verification: 9654adc879abd34218e7135dccbc9382f4b5623a

Scope

This audit covers the follow-up live validation of the Azure AI Inference endpoint after the earlier env-refresh retest had already confirmed repository checks, notebook execution, and a live Azure OpenAI chat-completions probe.

Executed Commands

Executed successfully in this turn:

  • set -a; source .env; set +a; uv run python - <<'PY' ... ChatCompletionsClient(...).complete(...) ... PY

Probe sequence:

  1. Read AZURE_INFERENCE_ENDPOINT, AZURE_INFERENCE_CREDENTIAL, AZURE_OPENAI_API_VERSION, AZURE_OPENAI_ENDPOINT, and AZURE_OPENAI_DEPLOYMENT_NAME from .env.
  2. Try the Azure AI Inference SDK against the current AZURE_INFERENCE_ENDPOINT value.
  3. Observe that the current env value is a full Azure OpenAI chat-completions URI and returns 404 Resource not found when passed directly to ChatCompletionsClient, because the SDK appends /chat/completions itself.
  4. Derive the SDK-ready base endpoint by stripping /chat/completions and the query string from the env value, then rerun the probe.

Notable Results

  • Direct SDK use against the current AZURE_INFERENCE_ENDPOINT env value failed with 404 Resource not found.
  • The derived base endpoint path /openai/deployments/gpt-4o succeeded with the same credential.
  • The live Azure AI Inference round trip returned:
  • reply: INFERENCE_OK
  • resolved model: gpt-4o-2024-11-20
  • finish reason: stop
  • No model argument was required once the deployment was encoded in the endpoint path.

Current Verification Status

Configured and verified in this turn:

  • Live Azure AI Inference endpoint validation: present and passed through the minimal ChatCompletionsClient.complete(...) probe, using a deployment-base endpoint derived from the current env value

Still absent or not exercised in this turn:

  • UI or browser tests: none found in the repository configuration
  • Automated DSPy training compile path: not implemented in the repository today

Notes

  • The current .env stores AZURE_INFERENCE_ENDPOINT as a full request URI ending in /chat/completions?.... That shape works for direct REST calls but not for the Azure AI Inference SDK constructor.
  • The repository documentation now clarifies that SDK callers should use the deployment base endpoint form instead.
  • The broader repository retest evidence remains in docs/audit/2026-03-18-retest-with-env-refresh.md.