feat: метрики качества — арифметика (sum==price*qty) + частота JSON-фиксов

deploy/services/classify.py

@@ -56,7 +56,7 @@ def _is_garbage_by_header(text: str) -> bool:
 def _call_llm_classify(header_text):
     """
     Прямой вызов LLM для классификации ОДНОГО документа.
-    Возвращает (parsed_dict, raw_text).
+    Возвращает (parsed_dict, raw_text, needed_fix).
     """
     prompt, _ = build_classify_prompt(header_text)
     payload = {
@@ -74,7 +74,8 @@ def _call_llm_classify(header_text):
         data = resp.json()
 
     raw = data.get("choices", [{}])[0].get("message", {}).get("content", "")
-    return _safe_json_parse(raw), raw
+    parsed, needed_fix = _safe_json_parse(raw)
+    return parsed, raw, needed_fix
 
 
 def classify_batch(batch_id):
@@ -93,23 +94,32 @@ def classify_batch(batch_id):
     total = len(pending)
     done = 0
     failed = 0
+    garbage = 0
+    json_fixes = 0
+    json_total = 0
 
     def _classify_one(doc):
         """Классифицировать один документ: фильтр → выжимка → LLM → сохранить."""
+        nonlocal garbage, json_fixes, json_total
         try:
             # Stage 1: garbage by filename (0 tokens)
             if _is_garbage_by_filename(doc["filename"]):
                 db_docs.set_classify_garbage(doc["id"], "filename_regex")
+                garbage += 1
                 return True
 
             # Stage 2: garbage by header keywords (0 tokens)
             text = _smart_extract(doc["elements_json"])
             if _is_garbage_by_header(text):
                 db_docs.set_classify_garbage(doc["id"], "header_keywords")
+                garbage += 1
                 return True
 
             # Stage 3: LLM classification (only for remaining)
-            result, raw = _call_llm_classify(text)
+            result, raw, needed_fix = _call_llm_classify(text)
+            json_total += 1
+            if needed_fix:
+                json_fixes += 1
             db_docs.set_classification(
                 doc["id"],
                 result.get("doc_type", "other"),
@@ -133,11 +143,13 @@ def classify_batch(batch_id):
             else:
                 failed += 1
 
-    return {"ok": True, "total": total, "done": done, "failed": failed}
+    return {"ok": True, "total": total, "done": done, "failed": failed,
+            "garbage": garbage, "json_fix_rate": round(json_fixes / max(json_total, 1), 3)}
 
 
 def _safe_json_parse(raw):
-    """Parse LLM response, fixing common JSON errors."""
+    """Parse LLM response, fixing common JSON errors.
+    Returns (parsed_dict, needed_fix: bool)."""
     if not raw:
         raise ValueError("empty LLM response")
 
@@ -156,7 +168,7 @@ def _safe_json_parse(raw):
 
     # Try strict parse
     try:
-        return json.loads(text)
+        return json.loads(text), False
     except json.JSONDecodeError:
         pass
 
@@ -169,7 +181,7 @@ def _safe_json_parse(raw):
 
     # Try again
     try:
-        return json.loads(text)
+        return json.loads(text), True
     except json.JSONDecodeError:
         pass
 
@@ -185,7 +197,7 @@ def _safe_json_parse(raw):
             text += '"'
         text += "}"
 
-    return json.loads(text)
+    return json.loads(text), True
 
 
 def _smart_extract(elements_json):