B2: tolerant JSON extraction from Claude responses + graceful fallback

_extract_json_payload handles a json/JSON/bare fence, raw JSON and JSON embedded in prose; any unparseable response or contract violation now degrades to the deterministic scoring_service instead of raising. Also guards the response envelope itself (content[0].text). B3: single home for recommendation scoring services/recommendation_service.py holds the rules; routers/stats.py and backend/services/stats_service.py both delegate to it. Unified rules are the union of the two old copies: same weights/threshold, substring matching (superset of the old exact match), tolerant key aliases, health_flags rule kept. Endpoint response contract unchanged. Plus: Overpass circuit breaker and concurrent zone queries in geo_service - 128 sequential calls per analysis no longer each burn a connect timeout when the host has no outbound network. Tests: 152 -> 194 passed.
2026-07-25 13:01:02 +00:00
parent 76278f5fe1
commit 8b3a2cbf7e
7 changed files with 509 additions and 59 deletions
@@ -171,21 +171,56 @@ async def analyze_with_claude(case_data: dict, api_key: str) -> AnalysisResult:
        if response.status_code != 200:
            raise Exception(f"Anthropic API error: {response.status_code} - {response.text}")

-        result = response.json()
-        content = result["content"][0]["text"]
+        try:
+            result = response.json()
+            content = result["content"][0]["text"]
+            analysis_data = _extract_json_payload(content)
+            analysis_data['fallback_used'] = False
+            return AnalysisResult(**analysis_data)
+        except Exception as e:
+            # Ответ модели пришёл в неожидаемом виде — не роняем анализ,
+            # а отдаём детерминированный результат scoring_service.
+            logger.warning(
+                f"Не удалось разобрать ответ Claude ({type(e).__name__}: {e}). "
+                "Используется fallback scoring service."
+            )
+            return await analyze_with_fallback(case_data)

-        # Парсим JSON из ответа
-        # Убираем возможные markdown блоки кода
-        if "```json" in content:
-            content = content.split("```json")[1].split("```")[0].strip()
-        elif "```" in content:
-            content = content.split("```")[1].split("```")[0].strip()

-        analysis_data = json.loads(content)
-        analysis_data['fallback_used'] = False
+def _extract_json_payload(content: str) -> dict:
+    """Extract a JSON object from a model response.

-        # Преобразуем в Pydantic модель
-        return AnalysisResult(**analysis_data)
+    Tolerates a ```json fence, a bare ``` fence, or raw JSON with
+    surrounding prose. Raises ValueError if nothing parseable is found.
+    """
+    candidates = []
+    stripped = (content or '').strip()
+
+    if '```' in stripped:
+        for marker in ('```json', '```JSON', '```'):
+            if marker in stripped:
+                after = stripped.split(marker, 1)[1]
+                candidates.append(after.split('```', 1)[0].strip())
+                break
+
+    candidates.append(stripped)
+
+    # Last resort: the widest {...} span in the text.
+    start, end = stripped.find('{'), stripped.rfind('}')
+    if start != -1 and end > start:
+        candidates.append(stripped[start:end + 1])
+
+    for candidate in candidates:
+        if not candidate:
+            continue
+        try:
+            parsed = json.loads(candidate)
+        except (json.JSONDecodeError, TypeError):
+            continue
+        if isinstance(parsed, dict):
+            return parsed
+
+    raise ValueError('Не удалось извлечь JSON из ответа модели')


 async def analyze_with_fallback(case_data: dict) -> AnalysisResult: