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("/")
+CHATBOT_BASE_URL = os.environ.get("DRE_CHATBOT_BASE_URL", "https://admin-ai.itpropartner.com/v1").rstrip("/")
-DEEPSEEK_API_KEY = os.environ.get("DEEPSEEK_API_KEY", "")
+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:
+def _call_llm(system_prompt: str, history: list[dict], message: str) -> str | None:
-    if not DEEPSEEK_API_KEY:
+    if not CHATBOT_API_KEY:
-        logger.error("DEEPSEEK_API_KEY not set; chatbot cannot call LLM")
+        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
+    # deepseek-v4-flash reasons even through LiteLLM; with max_tokens set high
-    # surface the model's reasoning; treat it as a failed call instead.
+    # 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}
 

backend/scripts/create-chatbot-key.sh

@@ -0,0 +1,58 @@
+#!/usr/bin/env bash
+# Idempotent: ensure a single dre-chatbot virtual key exists on admin-ai.
+# LiteLLM facts (learned 2026-08-26):
+#   - /key/list -> {"keys": [hash...], total_count, total_pages}  (paginated)
+#   - /key/info?key=<hash> -> {"key": "<hash>", "info": {key_alias, models, ...}}
+#   - /key/generate -> {"key": "sk-...", "token": "<hash>", "token_id": "<hash>"}
+#   - raw "sk-" key is ONLY visible in the /key/generate response; never stored
+#     in plaintext afterward (key_name is masked). So we must regenerate to get it.
+set -euo pipefail
+
+APP1_KEY=/root/.ssh/itpp-infra
+APP1_HOST=root@152.53.36.131
+ENV_FILE=/opt/dre-portal/.env
+ALIAS=dre-chatbot
+MODEL=deepseek-v4-flash
+BUDGET=5.0
+BASE="https://admin-ai.itpropartner.com"
+
+MK=$(ssh -i "$APP1_KEY" "$APP1_HOST" "grep LITELLM_MASTER_KEY /root/docker/litellm/.env | cut -d= -f2" 2>/dev/null)
+[ -n "$MK" ] || { echo "ERROR: no master key"; exit 1; }
+
+# --- 1. Delete any existing keys carrying the dre-chatbot alias (junk from bad runs)
+curl -s -H "Authorization: Bearer $MK" "$BASE/key/list" > /tmp/dre_keylist.json
+HASHES=$(python3 -c "import json; print('\n'.join(json.load(open('/tmp/dre_keylist.json')).get('keys', [])))")
+DEL=""
+for h in $HASHES; do
+  [ -z "$h" ] && continue
+  alias=$(curl -s -H "Authorization: Bearer $MK" "$BASE/key/info?key=$h" \
+    | python3 -c "import sys,json; print((json.load(sys.stdin).get('info') or {}).get('key_alias',''))" 2>/dev/null || echo "")
+  if [ "$alias" = "$ALIAS" ] || [ "$alias" = "dre-chatbot-inspect" ]; then
+    DEL="$DEL $h"
+  fi
+done
+if [ -n "$DEL" ]; then
+  curl -s -X POST "$BASE/key/delete" -H "Authorization: Bearer $MK" -H "Content-Type: application/json" \
+    -d "$(python3 -c "import json,sys; print(json.dumps({'keys': sys.argv[1].split()}))" "$DEL")" >/dev/null
+  echo "deleted stale keys:$DEL"
+fi
+
+# --- 2. Generate a fresh key, capture the RAW sk- token from the 'key' field
+RESP=$(curl -s -X POST "$BASE/key/generate" -H "Authorization: Bearer $MK" -H "Content-Type: application/json" \
+  -d "{\"key_alias\":\"$ALIAS\",\"models\":[\"$MODEL\"],\"max_budget\":$BUDGET,\"budget_duration\":\"daily\",\"metadata\":{\"purpose\":\"Agent DRE public FAQ chatbot\",\"service\":\"dre-chatbot\"}}")
+KEY=$(echo "$RESP" | python3 -c "import sys,json; print(json.load(sys.stdin).get('key',''))")
+case "$KEY" in
+  sk-*) ;;
+  *) echo "ERROR: generated key not sk- (got '${KEY:0:8}')"; echo "$RESP" | head -c 500; exit 1;;
+esac
+
+# --- 3. Write to .env (never print the key)
+if grep -q '^DRE_CHATBOT_API_KEY=' "$ENV_FILE"; then
+  sed -i "s|^DRE_CHATBOT_API_KEY=.*|DRE_CHATBOT_API_KEY=$KEY|" "$ENV_FILE"
+else
+  echo "DRE_CHATBOT_API_KEY=$KEY" >> "$ENV_FILE"
+fi
+grep -q '^DRE_CHATBOT_BASE_URL=' "$ENV_FILE" || echo "DRE_CHATBOT_BASE_URL=https://admin-ai.itpropartner.com/v1" >> "$ENV_FILE"
+grep -q '^DRE_CHATBOT_MODEL=' "$ENV_FILE" || echo "DRE_CHATBOT_MODEL=$MODEL" >> "$ENV_FILE"
+
+echo "DONE: alias=$ALIAS model=$MODEL budget=\$$BUDGET/day keylen=${#KEY}"