feat: migrate LLM backend from HF InferenceAPI to Cerebras

- Replace huggingface_hub InferenceClient → openai SDK (Cerebras-compatible)
- API endpoint: https://api.cerebras.ai/v1
- New models: llama3.3-70b, qwen-3-32b, llama3.1-8b (~1s each, free tier)
- Env var: CEREBRAS_API_KEY (was HF_TOKEN)
- requirements: huggingface_hub → openai>=1.30.0
- Expected speed: ~2100 tok/s vs ~100 tok/s on HF free tier

app.py

@@ -1,7 +1,7 @@
 """
 HK UTM LLM Assistant — Hugging Face Spaces deployment
 ======================================================
-Pure RAG pipeline: FAISS + sentence-transformers + HF InferenceClient
+Pure RAG pipeline: FAISS + sentence-transformers + Cerebras Inference API
 No langchain dependency — avoids pydantic v1/v2 conflicts on Python 3.13.
 
 Features:
@@ -13,7 +13,7 @@ Features:
   - Eager pipeline load at startup (background thread)
 
 Environment variables (set as HF Secrets):
-  HF_TOKEN    : Your HuggingFace API token
+  CEREBRAS_API_KEY : Your Cerebras Cloud API key (csk-...)
   HF_MODEL_ID : (optional) defaults to Qwen/Qwen2.5-72B-Instruct
 """
 
@@ -25,7 +25,7 @@ import gradio as gr
 from pathlib import Path
 
 # ── Config ────────────────────────────────────────────────────────────────────
-HF_TOKEN    = os.environ.get("HF_TOKEN", "")
+CEREBRAS_API_KEY = os.environ.get("CEREBRAS_API_KEY", "")
 INDEX_DIR   = "data/processed/faiss_index"
 DATA_DIR    = "data/raw"
 EMBED_MODEL = "sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2"
@@ -41,25 +41,22 @@ CE_POOL      = 6   # reduced from 12 → faster reranking, RRF pre-filters so qu
 CE_MODEL     = "cross-encoder/ms-marco-MiniLM-L-6-v2"
 
 # ── Available LLM models (all free via HF Inference API) ─────────────────────
+# Cerebras Inference API — OpenAI-compatible, ~2100 tok/s, 1M tokens/day free
 LLM_MODELS = {
-    "Qwen2.5-72B (Default ★)": {
-        "id":   "Qwen/Qwen2.5-72B-Instruct",
-        "desc": "Best quality · strongest Chinese/English reasoning · ~20s",
+    "Llama-3.3-70B (Default ★)": {
+        "id":   "llama3.3-70b",
+        "desc": "最佳質素 · 強大中英文推理 · Cerebras ~1s",
     },
-    "Llama-3.3-70B": {
-        "id":   "meta-llama/Llama-3.3-70B-Instruct",
-        "desc": "Meta flagship · excellent English · fast ~5s",
+    "Qwen3-32B": {
+        "id":   "qwen-3-32b",
+        "desc": "Qwen3 旗艦 · 思維模式 · 強中文 · ~1s",
     },
-    "Qwen3-8B (Fast)": {
-        "id":   "Qwen/Qwen3-8B",
-        "desc": "Lightweight · thinking-mode capable · ~3s",
-    },
-    "Qwen2.5-7B (Fastest)": {
-        "id":   "Qwen/Qwen2.5-7B-Instruct",
-        "desc": "Fastest response · good for quick checks · ~2s",
+    "Llama-3.1-8B (Fast)": {
+        "id":   "llama3.1-8b",
+        "desc": "輕量快速 · 適合快速測試 · <1s",
     },
 }
-DEFAULT_MODEL_NAME = "Qwen2.5-72B (Default ★)"
+DEFAULT_MODEL_NAME = "Llama-3.3-70B (Default ★)"
 MODEL_NAMES        = list(LLM_MODELS.keys())
 
 UTM_SYSTEM_PROMPT = """You are an expert assistant in UAS Traffic Management (UTM) \
@@ -155,7 +152,7 @@ def _load_pipeline():
     try:
         import faiss
         from sentence_transformers import SentenceTransformer
-        from huggingface_hub import InferenceClient
+        from openai import OpenAI  # Cerebras is OpenAI-compatible
 
         print("=== Loading pipeline at startup ===")
         print("Loading embedding model...")
@@ -188,11 +185,13 @@ def _load_pipeline():
         print("Cross-encoder (mxbai-rerank-large-v2) loaded.")
 
         print("Initialising LLM clients...")
-        llm_clients = {
-            name: InferenceClient(model=cfg["id"], token=HF_TOKEN)
-            for name, cfg in LLM_MODELS.items()
-        }
-        print(f"LLM clients ready: {list(llm_clients.keys())}")
+        # Single Cerebras client — model is passed per-call
+        cerebras_client = OpenAI(
+            api_key=CEREBRAS_API_KEY,
+            base_url="https://api.cerebras.ai/v1",
+        )
+        llm_clients = {name: cerebras_client for name in LLM_MODELS}
+        print(f"Cerebras client ready: {list(llm_clients.keys())}")
 
         _pipeline = {
             "embed_model":   embed_model,
@@ -375,8 +374,11 @@ def chat(user_message: str, history: list, sidebar_state: str, model_name: str =
 
     # Non-streaming single call — more stable on mobile / weak connections
     try:
-        response = llm_client.chat_completion(
-            messages=messages, max_tokens=1024, temperature=0.3, stream=False
+        response = llm_client.chat.completions.create(
+            model=LLM_MODELS[model_name]["id"],
+            messages=messages,
+            max_tokens=1024,
+            temperature=0.3,
         )
         answer = response.choices[0].message.content or ""
     except Exception as e:

requirements.txt

@@ -1,4 +1,4 @@
-huggingface_hub>=0.23.0
+openai>=1.30.0          # Cerebras OpenAI-compatible API
 faiss-cpu>=1.7.4
 sentence-transformers>=3.0.0,<4.0.0
 rank_bm25>=0.2.2