new model

app.py

@@ -179,18 +179,12 @@ qa_prompt = ChatPromptTemplate.from_template(
 # =============================
 # 4) LLM — local, token-free
 # =============================
-LOCAL_MODEL_ID = os.environ.get("LOCAL_LLM", "mistralai/Mixtral-8x7B-v0.1")
-tok = AutoTokenizer.from_pretrained(LOCAL_MODEL_ID)
-mdl = AutoModelForSeq2SeqLM.from_pretrained(LOCAL_MODEL_ID)
-
-gen = pipeline(
-    task="text2text-generation",
-    model=mdl,
-    tokenizer=tok,
-    do_sample=False,  # deterministic; helps JSON adherence
-)
-llm = HuggingFacePipeline(pipeline=gen)
-
+llm = HuggingFaceHub( repo_id="mistralai/Mixtral-8x7B-v0.1", 
+                     task="text-generation", 
+                     model_kwargs={ "max_new_tokens": 2000, 
+                                   "top_k": 30, 
+                                   "temperature": 0.1, 
+                                   "repetition_penalty": 1.03 }, )
 
 # ===========================================
 # 5) Chain (memory + robust JSON extraction)