[SagiEv]

Summary

Fixes AI Role Fit Analysis (POST /role-fit/analyze) failing with 504 DEADLINE_EXCEEDED / 500 Internal Server Error and falling back solely to the deterministic score.

Root Cause

1. Payload Bloat (backend/services/fitAnalysis.service.js): getFitContext passed raw SELECT * rows from skills, profile, and user_experiences directly into candidate_data. Each skill included its 384-dimensional float embedding vector (vector(384)), inflating json.dumps(request.candidate_data) in the LLM prompt by tens of thousands of tokens.
2. Aggressive Gemini Timeout (backend/ai_service/providers/gemini_provider.py): A 12-second server deadline (timeout=12) caused ChatGoogleGenerativeAI structured-output requests to time out with 504 DEADLINE_EXCEEDED across primary and fallback models.

Changes

• Node Backend (backend/services/fitAnalysis.service.js):
  • Strip embedding vectors and internal DB metadata from skills, experiences, and profile in getFitContext.
  • Pass configured fitConfig.model to /role-fit/analyze, bound job_description length (15000 chars), and add default AI_SERVICE_URL fallback.
• Python AI Service (backend/ai_service/providers/gemini_provider.py, backend/ai_service/role_fit/):
  • Increase ChatGoogleGenerativeAI timeout from 12s to 45s.
  • Omit temperature parameter for gemini-3.5-flash-lite to avoid fixed-sampling warnings.
  • Accept optional model in FitAnalysisRequest and forward it to get_fast_llm.
• Frontend (frontend/src/pages/AnalyticsPage.jsx):
  • Read fit_analysis_ai?.ai_score before falling back to overall_score and fit_score_deterministic.
• Tests (backend/tests/fitAnalysis.test.js):
  • Added unit test verifying getFitContext strips skill embeddings and internal metadata.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
job-pilot	Ready	Preview	Nov 5, 2026 7:53pm UTC