v1.0.175: виртуальные группы + classify_raw + stepper

deploy/services/classify.py

@@ -31,15 +31,14 @@ LLM_MODEL = "gpt-oss-120b"
 def _call_llm_classify(header_text):
     """
     Прямой вызов LLM для классификации ОДНОГО документа.
-    Не использует общий call_llm() из services/llm.py — здесь свой промпт и формат ответа.
-    Возвращает распарсенный JSON dict с полями: doc_type, own_number, parent_number, doc_date, counterparty, confidence.
+    Возвращает (parsed_dict, raw_text).
     """
     prompt, _ = build_classify_prompt(header_text)
     payload = {
         "model": LLM_MODEL,
         "messages": [{"role": "user", "content": prompt}],
-        "max_tokens": 1000,  # достаточно для JSON с длинными названиями
-        "temperature": 0.1,  # минимальная температура для детерминированности
+        "max_tokens": 1000,
+        "temperature": 0.1,
     }
     with httpx.Client(http2=True, timeout=60, verify=False) as client:
         resp = client.post(
@@ -50,7 +49,7 @@ def _call_llm_classify(header_text):
         data = resp.json()
 
     raw = data.get("choices", [{}])[0].get("message", {}).get("content", "")
-    return _safe_json_parse(raw)
+    return _safe_json_parse(raw), raw
 
 
 def classify_batch(batch_id):
@@ -74,7 +73,7 @@ def classify_batch(batch_id):
         """Классифицировать один документ: выжимка → LLM → сохранить результат."""
         try:
             text = _smart_extract(doc["elements_json"])
-            result = _call_llm_classify(text)
+            result, raw = _call_llm_classify(text)
             db_docs.set_classification(
                 doc["id"],
                 result.get("doc_type", "other"),
@@ -82,6 +81,7 @@ def classify_batch(batch_id):
                 result.get("parent_number"),
                 result.get("doc_date"),
                 result.get("counterparty"),
+                classify_raw=raw,
             )
             return True
         except Exception as e:

deploy/services/grouping.py

@@ -33,8 +33,9 @@ def group_documents(batch_id):
     1. Отделить contract от supplement/specification
     2. Каждый contract → якорь группы
     3. Для каждого supplement: найти contract по parent_number (нормализованный)
-    4. Несматченные → группа "__unresolved__"
-    5. Внутри группы сортировка по doc_date
+    4. Оставшиеся supplement/spec → виртуальные группы по parent_number/own_number
+    5. Совсем без номеров → группа "__unresolved__"
+    6. Внутри группы сортировка по doc_date
 
     Возвращает {ok, groups: [{contract_number, counterparty, documents: [...]}], total_docs}.
     """
@@ -79,18 +80,43 @@ def group_documents(batch_id):
 
         groups.append(group)
 
-    # Unmatched documents → "unresolved" group (including failed/pending)
-    unmatched = [s for s in supplements_list if s.get("doc_type") != "contract"]
+    # ── Virtual groups: unmatched supplements grouped by number ──────────
+    # Group remaining supplements/specs by normalized parent_number (priority) or own_number
+    virtual = {}
+    for s in supplements_list:
+        num = normalize_number(s.get("parent_number") or s.get("own_number") or "")
+        if not num:
+            continue  # no number → stays in supplements_list for unresolved
+        if num not in virtual:
+            # Use counterparty from first doc in group
+            cp = s.get("counterparty") or ""
+            virtual[num] = {
+                "contract_number": s.get("parent_number") or s.get("own_number") or "?",
+                "counterparty": cp,
+                "documents": [],
+            }
+        virtual[num]["documents"].append(s)
+        # Update counterparty if current doc has a better one
+        if not virtual[num]["counterparty"] and s.get("counterparty"):
+            virtual[num]["counterparty"] = s.get("counterparty")
+
+    for vnum, vgroup in virtual.items():
+        vgroup["documents"].sort(key=lambda x: x.get("doc_date") or "")
+        groups.append(vgroup)
+        # Remove grouped docs from supplements_list
+        for s in vgroup["documents"]:
+            supplements_list.remove(s)
+
+    # ── Unresolved: everything left (no number, failed, pending, other) ──
+    unmatched = [s for s in supplements_list]  # remaining after virtual grouping
     unmatched += [d for d in docs if d.get("classify_status") != "classified"]
 
-    if unmatched or len(contracts_list) == 0:
-        # Put unmatched into a separate group
-        if unmatched:
-            groups.append({
-                "contract_number": "__unresolved__",
-                "counterparty": "",
-                "documents": unmatched,
-            })
+    if unmatched:
+        groups.append({
+            "contract_number": "__unresolved__",
+            "counterparty": "",
+            "documents": unmatched,
+        })
 
     return {"ok": True, "groups": groups, "total_docs": len(docs)}