feat(Ф2): llm_client.py — DI LLM (HttpxLLMClient + FakeLLMClient)

deploy/services/classify.py

@@ -27,6 +27,17 @@ 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"
 
+# Ленивый singleton — обратная совместимость
+_classify_llm = None
+
+
+def _get_classify_client():
+    global _classify_llm
+    if _classify_llm is None:
+        from .llm_client import HttpxLLMClient
+        _classify_llm = HttpxLLMClient(url=LLM_URL, key=LLM_KEY, model=LLM_MODEL, max_tokens=1000, timeout=60)
+    return _classify_llm
+
 # ── 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)',
@@ -53,29 +64,19 @@ def _is_garbage_by_header(text: str) -> bool:
     return any(marker in header for marker in _GARBAGE_HEADER_MARKERS)
 
 
-def _call_llm_classify(header_text):
+def _call_llm_classify(header_text, llm_client=None):
     """
     Прямой вызов LLM для классификации ОДНОГО документа.
     Возвращает (parsed_dict, raw_text, needed_fix).
+    llm_client: LLMClient (optional). Default — HttpxLLMClient.
     """
-    prompt, _ = build_classify_prompt(header_text)
-    payload = {
-        "model": LLM_MODEL,
-        "messages": [{"role": "user", "content": prompt}],
-        "max_tokens": 1000,
-        "temperature": 0.1,
-    }
-    with httpx.Client(http2=True, timeout=60) as client:
-        resp = client.post(
-            LLM_URL, json=payload,
-            headers={"Authorization": f"Bearer {LLM_KEY}", "Content-Type": "application/json"},
-        )
-        resp.raise_for_status()
-        data = resp.json()
+    if llm_client is None:
+        llm_client = _get_classify_client()
 
-    raw = data.get("choices", [{}])[0].get("message", {}).get("content", "")
-    parsed, needed_fix = _safe_json_parse(raw)
-    return parsed, raw, needed_fix
+    prompt, _ = build_classify_prompt(header_text)
+    raw_text = llm_client.complete(prompt)
+    parsed, needed_fix = _safe_json_parse(raw_text)
+    return parsed, raw_text, needed_fix
 
 
 def classify_batch(batch_id):