cannot use with my desired GPT model

[kelvinho0215]

when i try to use Azure GPT-5-mini and Azure GPT-4.1-mini etc,

when try to ingestion the schema, it always using GPT4.1 and causing error. no matter i changed the environment model in my docker compose template

COMPLETION_MODEL=gpt-5-mini
EMBEDDING_MODEL=text-embedding-ada-002

[overcut-ai]

Completed Working on "Automated Ticket Triage "

✅ Triage complete for issue #572. Applied labels: bug. No duplicate issue identified from targeted top search results (no >=0.75 similarity match). needs-info was not applied because the report includes a concrete reproducible symptom and relevant environment variables.

✅ Workflow completed successfully.

---

👉 View complete log

[overcut-ai]

Completed Working on "Auto Root Cause Analysis"

✅ Workflow completed successfully.

---

👉 View complete log

[overcut-ai]

Root Cause Analysis (RCA)

Summary

The reported behavior is consistent with two code-level issues:

1. Graphiti Azure path hardcodes GPT-4.1, so schema/memory ingestion can ignore configured model overrides.
2. Azure completion override is not normalized to vendor/model format when user provides raw values (e.g., gpt-5-mini), which can cause provider/model mismatch behavior in parts of the pipeline.

Evidence

• Issue report: user sets COMPLETION_MODEL=gpt-5-mini and EMBEDDING_MODEL=text-embedding-ada-002, but ingestion still uses GPT-4.1.
• api/memory/graphiti_tool.py (around AzureOpenAIConfig.__init__):
  • os.environ["MODEL_NAME"] = "gpt-4.1"
  • self.model_choice = "gpt-4.1"
  • This directly forces GPT-4.1 regardless of env config.
• api/config.py (Azure branch):
  • COMPLETION_MODEL = _user_completion or "azure/gpt-4.1"
  • User-provided completion value is accepted as-is (no Azure prefix normalization), unlike other branches that normalize via helper logic.
• api/core/pipeline.py (validate_custom_model): expects vendor/model pattern for custom model values.
• README.md examples also use vendor/model format (azure/gpt-4.1, etc.).

Impacted Areas

• Azure-based schema/memory ingestion paths that pass through Graphiti config.
• Users overriding models with bare names (without provider prefix) may experience inconsistent behavior.

---

Minimal Safe Fix Suggestion

Fix (small scope)

• In api/memory/graphiti_tool.py, remove hardcoded GPT-4.1 and derive completion/embedding deployment names from Config.COMPLETION_MODEL and Config.EMBEDDING_MODEL_NAME (normalize by stripping optional provider prefix).

Pseudo-diff (minimal)

diff --git a/api/memory/graphiti_tool.py b/api/memory/graphiti_tool.py
@@ class AzureOpenAIConfig:
-        os.environ["MODEL_NAME"] = "gpt-4.1"
-        self.model_choice = "gpt-4.1"
+        self.model_choice = extract_embedding_model_name(Config.COMPLETION_MODEL)
@@
-        self.small_model = os.getenv('AZURE_SMALL_MODEL', 'gpt-4o-mini')
+        self.small_model = extract_embedding_model_name(
+            os.getenv('AZURE_SMALL_MODEL', self.model_choice)
+        )

(Optionally add one info log to print selected Graphiti models for easier debugging.)

Risk Assessment

• Risk: Low
• Reason: change is isolated to Azure Graphiti config path and removes a hardcoded override to align with existing env-driven config behavior.

Test Recommendations

1. Unit: verify AzureOpenAIConfig uses configured completion/embedding models from Config for both azure/gpt-5-mini and bare deployment name inputs.
2. Integration: initialize Graphiti memory path with Azure enabled and confirm it no longer uses gpt-4.1 unless explicitly configured.
3. Manual repro: run with COMPLETION_MODEL=gpt-5-mini (or azure/gpt-5-mini) and confirm ingestion logs/telemetry show configured model, not GPT-4.1.

---

Confidence: 0.88 (high)

If you want, I can also provide a follow-up PR-ready patch for this exact change.

[kelvinho0215]

would this have roadmap to fix ?

[overcut-ai]

Thanks for the follow-up — yes, this is fixable and we have a clear path.

Proposed roadmap

1. Immediate patch (short-term)

   • Remove the hardcoded gpt-4.1 in the Azure Graphiti path.
   • Make Graphiti read the configured model from COMPLETION_MODEL / EMBEDDING_MODEL consistently.

2. Validation pass

   • Add focused tests to confirm ingestion uses the configured model (including Azure-prefixed and non-prefixed values).
   • Reproduce your scenario in a Docker setup before release.

3. Hardening (follow-up)

   • Improve model-name normalization/validation so invalid format is surfaced clearly.
   • Add clearer docs/examples for Azure model/deployment naming.

ETA expectation

• Patch + validation: usually suitable for a near-term release/hotfix cycle.
• Hardening/docs improvements: can follow right after in the next iteration.

If helpful, I can also draft the exact environment values to use once the patch is merged so you can verify quickly.