Update app/services/classifier.py

app/services/classifier.py

@@ -1,6 +1,7 @@
 from __future__ import annotations
 
 import json
+import logging
 import os
 import re
 from pathlib import Path
@@ -15,8 +16,7 @@ NSFW_WEIGHTS_PATH = MODELS_DIR / "model.safetensors"
 NSFW_THRESHOLD = 0.75
 
 LLM_MODELS = [
-    os.getenv("GAME_LLM_MODEL", "Qwen/Qwen2.5-7B-Instruct"),
-    "microsoft/Phi-3.5-mini-instruct",
+    os.getenv("GAME_LLM_MODEL", "Qwen/Qwen2.5-1.5B-Instruct"),
 ]
 LLM_MAX_CHARS = 3000
 
@@ -25,6 +25,8 @@ _nsfw_error: str | None = None
 _ocr_reader: Any | None = None
 _ocr_error: str | None = None
 
+logger = logging.getLogger(__name__)
+
 GAME_KEYWORDS = {
     "valorant",
     "steam",
@@ -118,8 +120,22 @@ def classify_screenshot(file_path: str, filename: str, suspected_game: bool) ->
 
 def classify_game_with_ocr_llm(file_path: str, filename: str, suspected_game: bool) -> dict[str, Any]:
     ocr_text, urls = extract_ocr_text_and_urls(file_path)
+    logger.info(
+        "game-detect start file=%s suspected_game=%s ocr_len=%d urls=%d",
+        filename,
+        suspected_game,
+        len(ocr_text),
+        len(urls),
+    )
     llm = _classify_game_with_llm(ocr_text, urls, suspected_game)
     if llm is not None:
+        logger.info(
+            "game-detect llm verdict=%s confidence=%.2f reason=%s source=%s",
+            llm["verdict"],
+            float(llm["confidence"]),
+            llm["reason"],
+            llm.get("source", "llm"),
+        )
         return {
             "verdict": llm["verdict"],
             "confidence": llm["confidence"],
@@ -134,6 +150,11 @@ def classify_game_with_ocr_llm(file_path: str, filename: str, suspected_game: bo
     keyword_hit = any(word in text_lower for word in GAME_KEYWORDS)
     domain_hit = any(_is_game_like_domain(url) for url in urls)
     if domain_hit or (suspected_game and keyword_hit):
+        logger.info(
+            "game-detect fallback verdict=game reason=ocr-keyword-heuristic domain_hit=%s keyword_hit=%s",
+            domain_hit,
+            keyword_hit,
+        )
         return {
             "verdict": "game",
             "confidence": 0.78,
@@ -144,6 +165,9 @@ def classify_game_with_ocr_llm(file_path: str, filename: str, suspected_game: bo
         }
 
     if suspected_game:
+        logger.info(
+            "game-detect fallback verdict=uncertain reason=signal-without-clear-ocr"
+        )
         return {
             "verdict": "uncertain",
             "confidence": 0.55,
@@ -153,6 +177,7 @@ def classify_game_with_ocr_llm(file_path: str, filename: str, suspected_game: bo
             "source": "heuristic",
         }
 
+    logger.info("game-detect fallback verdict=not_game reason=no-game-evidence")
     return {
         "verdict": "not_game",
         "confidence": 0.2,
@@ -166,12 +191,14 @@ def classify_game_with_ocr_llm(file_path: str, filename: str, suspected_game: bo
 def extract_ocr_text_and_urls(file_path: str) -> tuple[str, list[str]]:
     reader = _load_ocr_reader()
     if reader is None:
+        logger.warning("ocr unavailable: reader is None")
         return "", []
 
     try:
         results = reader.readtext(file_path, detail=0, paragraph=True)
         ocr_text = "\n".join(str(x) for x in results).strip()
     except Exception:
+        logger.exception("ocr read failed for file=%s", file_path)
         return "", []
 
     raw_urls = URL_REGEX.findall(ocr_text)
@@ -193,6 +220,8 @@ def extract_ocr_text_and_urls(file_path: str) -> tuple[str, list[str]]:
         seen.add(item)
         urls.append(item)
 
+    logger.info("ocr extracted text_len=%d urls=%d", len(ocr_text), len(urls))
+
     return ocr_text[:LLM_MAX_CHARS], urls
 
 
@@ -209,9 +238,11 @@ def _load_ocr_reader() -> Any | None:
         import easyocr
 
         _ocr_reader = easyocr.Reader(["en"], gpu=False)
+        logger.info("ocr reader initialized")
         return _ocr_reader
     except Exception as exc:
         _ocr_error = str(exc)
+        logger.warning("ocr reader init failed: %s", _ocr_error)
         return None
 
 
@@ -238,6 +269,7 @@ def _classify_game_with_llm(ocr_text: str, urls: list[str], suspected_game: bool
 
     for model_name in LLM_MODELS:
         try:
+            logger.info("llm try model=%s", model_name)
             client = InferenceClient(model=model_name, token=token or None)
             response = client.chat_completion(
                 messages=prompt_messages,
@@ -250,12 +282,15 @@ def _classify_game_with_llm(ocr_text: str, urls: list[str], suspected_game: bool
                 content = response.choices[0].message.content or ""
             parsed = _parse_llm_json(content)
             if parsed is None:
+                logger.warning("llm invalid json model=%s content=%s", model_name, content[:300])
                 continue
             parsed["source"] = f"llm:{model_name}"
             return parsed
-        except Exception:
+        except Exception as exc:
+            logger.warning("llm error model=%s err=%s", model_name, repr(exc))
             continue
 
+    logger.warning("llm all models failed; fallback to heuristic")
     return None