feat: agent_worker DeepSeek fallback & Sprint 003 spec

backend/agent_worker.py

@@ -13,8 +13,12 @@ import db
 
 log = logging.getLogger(__name__)
 
-QWEN_BASE_URL = os.environ.get("QWEN_BASE_URL", "http://localhost:8080")
+# LLM config: use local Qwen if QWEN_BASE_URL set, otherwise fall back to DeepSeek
+QWEN_BASE_URL = os.environ.get("QWEN_BASE_URL", "")
 QWEN_MODEL = os.environ.get("QWEN_MODEL", "current")
+DEEPSEEK_API_KEY = os.environ.get("DEEPSEEK_API_KEY", "")
+DEEPSEEK_BASE_URL = "https://api.deepseek.com"
+DEEPSEEK_MODEL = os.environ.get("DEEPSEEK_MODEL", "deepseek-chat")
 POLL_INTERVAL = int(os.environ.get("AGENT_POLL_INTERVAL", "30"))
 MAX_CONCURRENT = int(os.environ.get("AGENT_MAX_CONCURRENT", "3"))
 SYSTEM_PROMPT = os.environ.get(
@@ -27,8 +31,8 @@ SYSTEM_PROMPT = os.environ.get(
 semaphore = asyncio.Semaphore(MAX_CONCURRENT)
 
 
-async def _call_qwen(prompt: str, video_title: str, video_context: str) -> str:
-    """Call the local Qwen27B API with video context + user prompt."""
+async def _call_llm(prompt: str, video_title: str, video_context: str) -> str:
+    """Call local Qwen or DeepSeek API with video context + user prompt."""
     messages = [
         {"role": "system", "content": SYSTEM_PROMPT},
         {
@@ -45,11 +49,27 @@ async def _call_qwen(prompt: str, video_title: str, video_context: str) -> str:
     https_proxy = os.environ.get("HTTPS_PROXY") or os.environ.get("https_proxy")
     proxies = http_proxy or https_proxy
 
+    if QWEN_BASE_URL:
+        # Use local Qwen
+        base_url = QWEN_BASE_URL
+        model = QWEN_MODEL
+        headers = {}
+        log.debug("Using local Qwen at %s", base_url)
+    elif DEEPSEEK_API_KEY:
+        # Fall back to DeepSeek cloud
+        base_url = DEEPSEEK_BASE_URL
+        model = DEEPSEEK_MODEL
+        headers = {"Authorization": f"Bearer {DEEPSEEK_API_KEY}"}
+        log.debug("Using DeepSeek cloud API")
+    else:
+        raise RuntimeError("No LLM configured: set QWEN_BASE_URL or DEEPSEEK_API_KEY")
+
     async with httpx.AsyncClient(proxy=proxies, timeout=180) as client:
         resp = await client.post(
-            f"{QWEN_BASE_URL}/v1/chat/completions",
+            f"{base_url}/v1/chat/completions",
+            headers=headers,
             json={
-                "model": QWEN_MODEL,
+                "model": model,
                 "messages": messages,
                 "max_tokens": 2048,
                 "temperature": 0.3,
@@ -60,6 +80,10 @@ async def _call_qwen(prompt: str, video_title: str, video_context: str) -> str:
         return data["choices"][0]["message"]["content"]
 
 
+# Backward compat alias
+_call_qwen = _call_llm
+
+
 async def _process_item(item: dict) -> None:
     """Process a single agent_queue item."""
     video_id = item["video_id"]
@@ -102,7 +126,7 @@ async def _process_item(item: dict) -> None:
     video_context = "\n\n".join(context_parts) if context_parts else "No additional context available."
 
     try:
-        result = await _call_qwen(user_prompt, video["title"], video_context)
+        result = await _call_llm(user_prompt, video["title"], video_context)
         status = "done"
     except Exception as e:
         log.warning(f"Qwen API call failed for item {item['id']}: {e}")