Upload app.py

app.py

@@ -7,6 +7,8 @@ from __future__ import annotations
 import json
 import os
 import re
+import logging
+import time
 from pathlib import Path
 
 import gradio as gr
@@ -17,6 +19,8 @@ from pydantic import BaseModel, ConfigDict, Field, ValidationError
 from sentence_transformers import SentenceTransformer
 from retrieval_zerogpu import OriginRetrieval, MODEL_ID
 
+logger = logging.getLogger('origin.gemini')
+
 ROOT = Path(__file__).resolve().parent
 DATA = ROOT / 'assets/projects_clean.csv'
 VECTORS = ROOT / 'assets/project_vectors.npz'
@@ -129,22 +133,86 @@ def enhance(title: str, description: str, language: str, search_payload: dict):
     request = {'proposed_idea':{'title':idea.title,'description':idea.description,'language':lang},
                'retrieved_projects':context}
     model_id = os.getenv('ORIGIN_GEMINI_MODEL', 'gemini-2.5-flash')
+    # Use Gemini's supported Pydantic structured-output mode. For Gemini 2.5 Flash,
+    # disabling thinking prevents a large hidden token budget consuming the output
+    # budget, which can otherwise produce empty/truncated JSON responses.
+    config_kwargs = dict(
+        system_instruction=SYSTEM_INSTRUCTIONS,
+        response_mime_type='application/json',
+        response_schema=GeneratedAnalysis,
+        temperature=0.2,
+        max_output_tokens=4096,
+    )
+    if model_id.startswith('gemini-2.5-flash'):
+        config_kwargs['thinking_config'] = types.ThinkingConfig(thinking_budget=0)
+
     try:
-        client = genai.Client(api_key=key, http_options=types.HttpOptions(timeout=60000,
-               retry_options=types.HttpRetryOptions(attempts=1)))
-        response = client.models.generate_content(
-            model=model_id, contents=json.dumps(request, ensure_ascii=False),
-            config=types.GenerateContentConfig(
-                system_instruction=SYSTEM_INSTRUCTIONS,
-                response_mime_type='application/json',
-                response_json_schema=GeneratedAnalysis.model_json_schema(),
-                temperature=0.2, max_output_tokens=4096))
-        if not response.text:
-            raise ValueError('empty model response')
-        generated = GeneratedAnalysis.model_validate_json(response.text)
+        # 60 seconds per request; avoid SDK retries that can consume extra quota.
+        client = genai.Client(
+            api_key=key,
+            http_options=types.HttpOptions(
+                timeout=60000,
+                retry_options=types.HttpRetryOptions(attempts=1),
+            ),
+        )
+        for attempt in range(2):
+            try:
+                response = client.models.generate_content(
+                    model=model_id,
+                    contents=json.dumps(request, ensure_ascii=False),
+                    config=types.GenerateContentConfig(**config_kwargs),
+                )
+                break
+            except Exception as exc:
+                code = getattr(exc, 'code', None)
+                # Retry only a temporarily unavailable server; NEVER retry
+                # invalid API keys, bad requests, or exhausted Gemini quota.
+                if attempt == 0 and code in (502, 503, 504):
+                    logger.warning('origin.gemini_retry code=%s model=%s', code, model_id)
+                    time.sleep(2)
+                    continue
+                raise
+
+        # The SDK can return a validated Pydantic instance directly.
+        if isinstance(getattr(response, 'parsed', None), GeneratedAnalysis):
+            generated = response.parsed
+        elif getattr(response, 'text', None):
+            generated = GeneratedAnalysis.model_validate_json(response.text)
+        else:
+            # Log only finish-reason name; never log candidate content.
+            candidates = getattr(response, 'candidates', None) or []
+            reason = getattr(candidates[0], 'finish_reason', None) if candidates else None
+            logger.warning('origin.gemini_empty_response finish_reason=%s model=%s',
+                           getattr(reason, 'name', str(reason)) if reason else 'unknown', model_id)
+            raise ValueError('empty_gemini_response')
     except Exception as exc:
-        # Avoid exposing prompts, SDK exceptions or secrets to public users.
-        raise gr.Error('Gemini analysis unavailable or invalid. Please retry later.') from None
+        # Gemini SDK exceptions often contain private request details. Log only
+        # safe status and exception TYPE, never exception message or prompt.
+        code = getattr(exc, 'code', None)
+        if code == 429:
+            category = 'quota_or_rate_limit'
+        elif code in (401, 403):
+            category = 'credentials_or_permissions'
+        elif code == 404:
+            category = 'model_not_found'
+        elif code == 400:
+            category = 'invalid_request_or_schema'
+        elif code in (500, 502, 503, 504):
+            category = 'temporary_api_failure'
+        elif isinstance(exc, ValidationError):
+            category = 'generated_json_validation'
+        elif isinstance(exc, ValueError):
+            category = 'empty_or_invalid_json'
+        else:
+            category = 'sdk_or_network_error'
+        logger.error(
+            'origin.gemini_failed category=%s status=%s type=%s model=%s',
+            category, code if isinstance(code, int) else 'n/a',
+            type(exc).__name__, model_id,
+        )
+        # /search remains independent, and the real React frontend already
+        # preserves search results when /enhance fails.
+        raise gr.Error('Gemini analysis unavailable. Search results are still available.') from None
 
     # Recreate response from vetted local CSV and validated generated fields.
     ranks = [{'rank':i+1, 'project_id':pid} for i,pid in enumerate(ids)]