Route Agent D.R.E chatbot through admin-ai (dedicated dre-chatbot virtual key, $5/day budget) instead of native DeepSeek key

backend/chatbot.py

@@ -1,6 +1,6 @@
 """Agent D.R.E chatbot — public FAQ assistant.
 
-Backed by DeepSeek (deepseek-v4-flash, thinking disabled) via native API.
+Backed by deepseek-v4-flash via the admin-ai LiteLLM proxy (dedicated virtual key, budget capped).
 Stdlib-only (urllib + json) so no new venv dependencies.
 Public endpoint: POST /api/chat
 
@@ -31,10 +31,10 @@ logger = logging.getLogger("dre.chatbot")
 
 router = APIRouter()
 
-DEEPSEEK_BASE_URL = os.environ.get("DEEPSEEK_BASE_URL", "https://api.deepseek.com").rstrip("/")
-DEEPSEEK_API_KEY = os.environ.get("DEEPSEEK_API_KEY", "")
+CHATBOT_BASE_URL = os.environ.get("DRE_CHATBOT_BASE_URL", "https://admin-ai.itpropartner.com/v1").rstrip("/")
+CHATBOT_API_KEY = os.environ.get("DRE_CHATBOT_API_KEY", "")
 CHATBOT_MODEL = os.environ.get("DRE_CHATBOT_MODEL", "deepseek-v4-flash")
-CHATBOT_MAX_TOKENS = int(os.environ.get("DRE_CHATBOT_MAX_TOKENS", "300"))
+CHATBOT_MAX_TOKENS = int(os.environ.get("DRE_CHATBOT_MAX_TOKENS", "400"))
 CHATBOT_TIMEOUT = float(os.environ.get("DRE_CHATBOT_TIMEOUT", "30"))
 
 MAX_MESSAGE_CHARS = 1000
@@ -97,9 +97,9 @@ def _is_private_topic(text: str) -> bool:
     return any(m in low for m in PRIVATE_TOPIC_MARKERS)
 
 
-def _call_deepseek(system_prompt: str, history: list[dict], message: str) -> str | None:
-    if not DEEPSEEK_API_KEY:
-        logger.error("DEEPSEEK_API_KEY not set; chatbot cannot call LLM")
+def _call_llm(system_prompt: str, history: list[dict], message: str) -> str | None:
+    if not CHATBOT_API_KEY:
+        logger.error("DRE_CHATBOT_API_KEY not set; chatbot cannot call LLM")
         return None
 
     messages = [{"role": "system", "content": system_prompt}]
@@ -115,15 +115,14 @@ def _call_deepseek(system_prompt: str, history: list[dict], message: str) -> str
         "messages": messages,
         "max_tokens": CHATBOT_MAX_TOKENS,
         "temperature": 0.3,
-        "thinking": {"type": "disabled"},
     }
-    url = f"{DEEPSEEK_BASE_URL}/v1/chat/completions"
+    url = f"{CHATBOT_BASE_URL}/chat/completions"
     req = urllib.request.Request(
         url,
         data=json.dumps(payload).encode("utf-8"),
         headers={
             "Content-Type": "application/json",
-            "Authorization": f"Bearer {DEEPSEEK_API_KEY}",
+            "Authorization": f"Bearer {CHATBOT_API_KEY}",
         },
         method="POST",
     )
@@ -131,20 +130,21 @@ def _call_deepseek(system_prompt: str, history: list[dict], message: str) -> str
         with urllib.request.urlopen(req, timeout=CHATBOT_TIMEOUT) as resp:
             data = json.loads(resp.read().decode("utf-8"))
     except (urllib.error.URLError, urllib.error.HTTPError, TimeoutError, OSError, json.JSONDecodeError) as exc:
-        logger.error("deepseek call failed: %s", exc)
+        logger.error("LLM call failed: %s", exc)
         return None
 
     try:
         content = data["choices"][0]["message"].get("content") or ""
         content = content.strip()
     except (KeyError, IndexError, TypeError):
-        logger.error("unexpected deepseek response shape: %s", data)
+        logger.error("unexpected LLM response shape: %s", data)
         return None
 
-    # If thinking was not actually disabled, content can come back empty. Do not
-    # surface the model's reasoning; treat it as a failed call instead.
+    # deepseek-v4-flash reasons even through LiteLLM; with max_tokens set high
+    # enough, content is still populated. If content is empty, never surface the
+    # model's reasoning_content (chain-of-thought); treat it as a failed call.
     if not content:
-        logger.error("deepseek returned empty content (thinking may not be disabled)")
+        logger.error("LLM returned empty content")
         return None
     return content
 
@@ -193,7 +193,7 @@ async def chat(req: ChatRequest, request: Request):
         return {"reply": FALLBACK_OUT_OF_SCOPE}
 
     # 2. LLM answer from public FAQ only
-    reply = _call_deepseek(_system_prompt(), req.history, message)
+    reply = _call_llm(_system_prompt(), req.history, message)
     if reply is None:
         return {"reply": FALLBACK_ERROR}