Update app/text_to_text/farm_agent.py

app/text_to_text/farm_agent.py

@@ -16,7 +16,6 @@ from app.config import config
 log = logging.getLogger("farmlingua")
 
 SYSTEM_PROMPT = """You are FarmLingua AI, an expert agricultural assistant for Nigerian farmers built by Kawafarm LTD.
-
 RULES:
 - Answer ONLY farming-related questions — crops, livestock, soil, irrigation, fertilizer, pest control, harvest, and agribusiness.
 - If a question mentions a food item (e.g. "how to cook rice"), extract the crop name (rice) and answer how to PLANT and FARM it instead.
@@ -25,7 +24,6 @@ RULES:
 - Give complete, structured answers covering: land preparation, planting, inputs, care, harvest, and business or budget where relevant.
 - Never repeat the user's question. No preamble. Answer directly.
 - End every response by informing the user that Kawafarm LTD built you to help Nigerian farmers.
-
 FORMATTING RULES — CRITICAL:
 - Do NOT use markdown of any kind.
 - Do NOT use asterisks, hashtags, bold, italic, or any special characters for formatting.
@@ -66,7 +64,6 @@ class FarmAgent:
         )
         self._trans_model.eval()
 
-        
         self._validate_language_codes()
 
         
@@ -77,7 +74,7 @@ class FarmAgent:
         )
         self._llm_model = AutoModelForCausalLM.from_pretrained(
             config.LLM_MODEL,
-            torch_dtype=torch.float16 if device == "cuda" else torch.float32,
+            torch_dtype="auto",
             device_map="auto",
             token=token,
         )
@@ -85,7 +82,6 @@ class FarmAgent:
         log.info("FarmAgent ready.")
 
     def _validate_language_codes(self):
-        """Validate FLORES-200 codes resolve to real token IDs at startup."""
         log.info("[VALIDATE] Checking NLLB FLORES-200 token IDs...")
         for lang, code in config.NLLB_CODES.items():
             try:
@@ -96,7 +92,7 @@ class FarmAgent:
             except Exception as e:
                 log.error(f"  {lang} ({code}) failed — {e}")
 
-    # Language detection 
+    
 
     def detect_language(self, text: str) -> tuple:
         clean       = text.strip().replace("\n", " ")
@@ -187,7 +183,7 @@ class FarmAgent:
         text = re.sub(r'\n{3,}', '\n\n', text)
         return text.strip()
 
-    #  LLM streaming 
+    # LLM streaming 
 
     def stream_response(self, history: list, english_message: str):
         log.info(f"[LLM] Sending to Qwen: {english_message[:120]}")
@@ -218,17 +214,13 @@ class FarmAgent:
             kwargs=dict(
                 **model_inputs,
                 streamer=streamer,
-                max_new_tokens=config.LLM_MAX_TOKENS,
-                temperature=config.LLM_TEMPERATURE,
-                do_sample=True,
-                repetition_penalty=1.5,
-                no_repeat_ngram_size=4,
+                max_new_tokens=512,
             ),
         ).start()
 
         return streamer
 
-    # Full pipeline
+    #  Full pipeline 
 
     def process(self, user_text: str, history: list) -> dict:
         detected_lang, confidence = self.detect_language(user_text)
@@ -252,5 +244,5 @@ class FarmAgent:
         }
 
 
-# Singleton
+
 farm_agent = FarmAgent()