- Create docs/archive/2025-12-31-vertex-ai-auth-fix.md - Update README.md with recent changes entry - Correct status: Gemini via Vertex AI (not Google AI Studio) 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
2.6 KiB
ChatGoogleGenerativeAI Vertex AI Auth Fix
Date: 2025-12-31 PR: #147 Related: PR #145 (LangChain provider migration)
Background
PR #145 migrated from ChatVertexAI (gRPC) to ChatGoogleGenerativeAI (REST) to fix silent hangs during model initialization. However, the migration only configured API key authentication, breaking Cloud Run deployments that use service account credentials.
Problem
After deploying PR #145, Cloud Run lyrics worker failed with:
🤖 Failed to initialize chat model: GOOGLE_API_KEY environment variable is required for Google/Gemini models
The code required GOOGLE_API_KEY even though:
- Cloud Run already has
GOOGLE_CLOUD_PROJECTset - Service account credentials are available via ADC (Application Default Credentials)
- Vertex AI IAM permissions were already configured
Root Cause
ChatGoogleGenerativeAI supports TWO authentication backends:
| Backend | Trigger | Auth Method |
|---|---|---|
| Vertex AI | project param set |
Service account / ADC |
| Google AI Studio | Only google_api_key |
API key |
PR #145 only passed google_api_key, forcing the Google AI Studio backend which doesn't support service accounts.
Solution
Pass project parameter when GOOGLE_CLOUD_PROJECT is set to trigger Vertex AI backend:
# Before (broken)
model = ChatGoogleGenerativeAI(model=model_name, google_api_key=api_key)
# After (works with service account)
model_kwargs = {"model": model_name, ...}
if project:
model_kwargs["project"] = project # Triggers Vertex AI backend
if api_key:
model_kwargs["google_api_key"] = api_key # Optional fallback
model = ChatGoogleGenerativeAI(**model_kwargs)
Key Insight
ChatGoogleGenerativeAI from langchain-google-genai is a unified client that supports both Vertex AI and Google AI Studio APIs. The project parameter is the key discriminator - when present, it uses Vertex AI with ADC auth.
Testing
Verified that on Cloud Run with GOOGLE_CLOUD_PROJECT=nomadkaraoke:
- No
GOOGLE_API_KEYrequired - Uses attached service account via ADC
- REST API (avoids gRPC hang issues from PR #145)
Files Changed
lyrics_transcriber_temp/.../providers/model_factory.py- Added project param logicdocs/LESSONS-LEARNED.md- Updated documentation
Timing Analysis Context
This fix was discovered during a timing analysis comparing local vs remote job performance. Key findings:
- Remote jobs had 38+ minute delays in lyrics worker
- Root cause: gRPC hangs (fixed by PR #145) + auth misconfiguration (fixed by this PR)
- With both fixes, remote processing should match local performance