From 909468b5b7d7883302385966d3bf34ad846e5d15 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?=E2=80=9CNaeel=E2=80=9D?= <tazet@narod.ru>
Date: Sat, 27 Jun 2026 13:16:30 +0400
Subject: [PATCH] =?UTF-8?q?feat:=20=D1=82=D1=80=D1=91=D1=85=D1=8D=D1=82?=
 =?UTF-8?q?=D0=B0=D0=BF=D0=BD=D1=8B=D0=B9=20=D1=84=D0=B8=D0=BB=D1=8C=D1=82?=
 =?UTF-8?q?=D1=80=20=D0=BC=D1=83=D1=81=D0=BE=D1=80=D0=B0=20=E2=80=94=20?=
 =?UTF-8?q?=D1=8D=D1=82=D0=B0=D0=BF=D1=8B=201-2=20(0=20=D1=82=D0=BE=D0=BA?=
 =?UTF-8?q?=D0=B5=D0=BD=D0=BE=D0=B2)=20=D0=B2=20classify.py?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 deploy/db/documents.py      |  8 ++++++++
 deploy/services/classify.py | 38 ++++++++++++++++++++++++++++++++++++-
 2 files changed, 45 insertions(+), 1 deletion(-)

diff --git a/deploy/db/documents.py b/deploy/db/documents.py
index 527b5fa..e6d5d33 100644
--- a/deploy/db/documents.py
+++ b/deploy/db/documents.py
@@ -53,6 +53,14 @@ def set_classify_failed(doc_id, error):
     )
 
 
+def set_classify_garbage(doc_id, reason=""):
+    """Mark document as garbage (Stage 1-2 filter, no LLM call)."""
+    return execute(
+        "UPDATE documents SET doc_type='garbage', classify_status='garbage', error_message=%s WHERE id=%s",
+        (f"garbage: {reason}", doc_id),
+    )
+
+
 def list_pending(batch_id):
     """Documents waiting for classification."""
     return query(
diff --git a/deploy/services/classify.py b/deploy/services/classify.py
index aec3f23..1262c4c 100644
--- a/deploy/services/classify.py
+++ b/deploy/services/classify.py
@@ -27,6 +27,31 @@ LLM_URL = "https://api.aillm.ru/v1/chat/completions"
 LLM_KEY = os.environ.get("LLM_KEY") or os.environ.get("LLM_API_KEY", "")
 LLM_MODEL = "gpt-oss-120b"
 
+# ── Garbage filter (Stage 1: filename regex) ────────────────────
+_GARBAGE_FILENAME_RE = re.compile(
+    r'(сч[её]т|акт|плат[её]ж|УПД|сверк|инвойс|invoice|payment|act)',
+    re.IGNORECASE,
+)
+
+# ── Garbage filter (Stage 2: header keywords) ───────────────────
+_GARBAGE_HEADER_MARKERS = [
+    'СЧЕТ-ФАКТУРА', 'СЧЕТ НА ОПЛАТУ', 'АКТ СВЕРКИ',
+    'АКТ ОКАЗАННЫХ УСЛУГ', 'АКТ ВЫПОЛНЕННЫХ РАБОТ',
+    'ПЛАТЁЖНОЕ ПОРУЧЕНИЕ', 'УНИВЕРСАЛЬНЫЙ ПЕРЕДАТОЧНЫЙ',
+    'УПД', 'ПЛАТЕЖНОЕ ПОРУЧЕНИЕ',
+]
+
+
+def _is_garbage_by_filename(filename: str) -> bool:
+    """Stage 1: regex по имени файла — быстро, 0 токенов."""
+    return bool(_GARBAGE_FILENAME_RE.search(filename))
+
+
+def _is_garbage_by_header(text: str) -> bool:
+    """Stage 2: ключевые слова в первых 2KB текста — быстро, 0 токенов."""
+    header = text[:2000].upper()
+    return any(marker in header for marker in _GARBAGE_HEADER_MARKERS)
+
 
 def _call_llm_classify(header_text):
     """
@@ -70,9 +95,20 @@ def classify_batch(batch_id):
     failed = 0
 
     def _classify_one(doc):
-        """Классифицировать один документ: выжимка → LLM → сохранить результат."""
+        """Классифицировать один документ: фильтр → выжимка → LLM → сохранить."""
         try:
+            # Stage 1: garbage by filename (0 tokens)
+            if _is_garbage_by_filename(doc["filename"]):
+                db_docs.set_classify_garbage(doc["id"], "filename_regex")
+                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")
+                return True
+
+            # Stage 3: LLM classification (only for remaining)
             result, raw = _call_llm_classify(text)
             db_docs.set_classification(
                 doc["id"],