summary

ai_med_extract/api/routes.py

@@ -1146,8 +1146,6 @@ You are an expert clinical AI assistant. Your task is to generate a patient summ
 
 ### DATA TO SUMMARIZE ###
 - PATIENT VISITS: {visits}
-- BASELINE: {baseline}
-- DELTAS: {delta_text}
 
 ### EXAMPLE OUTPUT FORMAT ###
 ## Clinical Assessment
@@ -1165,7 +1163,7 @@ You are an expert clinical AI assistant. Your task is to generate a patient summ
 - Review recent lab results when available.
 </s>
 <|user|>
-Generate the 4-section patient summary in the exact format specified above.
+Start summary.
 </s>
 <|assistant|>
 ## Clinical Assessment
@@ -1175,11 +1173,12 @@ Generate the 4-section patient summary in the exact format specified above.
                     try:
                         raw_summary = pipeline.generate(
                             full_prompt,
-                            max_tokens=2000,
-                            temperature=0.1,
-                            top_p=0.9,
-                            # max_loops=3  # Allow up to 3 loops to complete sections
-                        )
+                            max_tokens=800,  # Reduced from 2000. The summary should be concise.
+                            temperature=0.0, # Set to 0.0 for maximum determinism and often faster generation.
+                            top_p=1.0,       # Set to 1.0 to consider all tokens, often faster than sampling with top_p < 1.0.
+                            # top_k=50,        # Adding top_k can sometimes speed up sampling. 50 is a good balance.
+                            # repeat_penalty=1.1 # A small penalty can prevent loops and potentially speed up convergence.
+                            )
                         logger.info(f"GGUF raw summary length: {len(raw_summary)} chars")
                     except Exception as gen_error:
                         logger.error(f"GGUF generation failed: {gen_error}")