Update app.py

app.py

@@ -6,6 +6,15 @@ import gradio as gr
 from huggingface_hub import hf_hub_download
 from llama_cpp import Llama
 
+# -------------------- Config --------------------
+TOP_K        = 10          
+FINAL_TOP_N  = 1           
+MIN_CONF     = 0.14        
+CHUNK_LIMIT  = 360          
+MAX_TOKENS   = 96           
+TEMP         = 0.2
+QUALITY_LOG  = "quality_feedback.jsonl"   
+
 # -------------------- Load Dataset --------------------
 DATASET_PATH = "nbb_merged_full.json"
 with open(DATASET_PATH, "r", encoding="utf-8") as f:
@@ -25,7 +34,7 @@ def normalize_record(d):
     }
 
 DOCS = [normalize_record(x) for x in RAW_DATA if normalize_record(x)["content"]["lo"].strip()]
-assert DOCS, "Dataset ไม่มี content.lo"
+assert DOCS, "Dataset ບໍ່ມີ content.lo"
 
 CORPUS = [d["content"]["lo"] for d in DOCS]
 IDS = [d["id"] for d in DOCS]
@@ -49,9 +58,12 @@ MODEL_PATH = hf_hub_download(
 
 LLM = Llama(
     model_path=MODEL_PATH,
-    n_ctx=2048,
-    n_threads=4,
-    n_gpu_layers=32
+    n_ctx=2048,       
+    n_threads=8,
+    n_gpu_layers=128,  
+    n_batch=512,      
+    logits_all=False,
+    verbose=False
 )
 
 SYSTEM_RULES = """
@@ -73,9 +85,13 @@ FORMAT:
 """
 
 def build_prompt(question, hits):
-    ctx = "\n\n".join([f"[{h['id']}] {ID2DOC[h['id']]['content']['lo']}" for h in hits])
+    ctx = "\n\n".join([
+        truncate(ID2DOC[h['id']]['content']['lo'])
+        for h in hits[:FINAL_TOP_N]
+    ])
     return f"{SYSTEM_RULES}\n\nContext:\n{ctx}\n\nQuestion:\n{question}\n\nAnswer:"
 
+
 def smart_answer(message):
     hits = search(message, k=3)
     if not hits or hits[0]["score"] < 0.1:
@@ -92,7 +108,7 @@ def respond(message, history):
     return history
 
 with gr.Blocks() as demo:
-    gr.Markdown("## 🌾 Lao Chatbot (NBB)")
+    gr.Markdown("## ທົດລອງ RDB Chatbot")
     chatbot_ui = gr.Chatbot()
     msg = gr.Textbox(placeholder="ພິມຄຳຖາມບ່ອນນີ້...")
     msg.submit(respond, [msg, chatbot_ui], chatbot_ui)