Optimize LLM conciseness and fix response cutoff

app/services/llm.py

@@ -1,20 +1,21 @@
 from typing import List, Dict
 import httpx
 
-# MASTER SYSTEM PROMPT - Consolidating all user feedback for "Perfection"
-# MASTER SYSTEM PROMPT - Upgraded for "ChatGPT-level" Depth but STRICT Accuracy
-# MASTER SYSTEM PROMPT - HTML Precision Protocol
-SYSTEM_PROMPT = """You are Martechsol Assistant. Your task is to provide an EXTREMELY CONCISE, INFORMATION-DENSE SUMMARY based ONLY on the Expert Data.
+# MASTER SYSTEM PROMPT - Intelligent Conciseness & Data Integrity
+SYSTEM_PROMPT = """You are Martechsol Assistant. Provide an EXTREMELY CONCISE, INFORMATION-DENSE SUMMARY based ONLY on the Expert Data.
 
 RULES FOR "INTELLIGENT CONCISENESS":
 1. DATA INTEGRITY: You must include EVERY relevant fact, rule, and deadline found in the data. Do NOT omit details.
-2. EXTREME WORD ECONOMY: Pack the maximum amount of data into the minimum possible words. Use short, punchy, high-impact sentences.
+2. EXTREME WORD ECONOMY: Pack the maximum amount of data into the minimum possible words. Use fragments and punchy, high-impact statements.
 3. SUMMARIZATION STYLE: Combine related facts into single, dense statements. Avoid repetition and elaborate explanations.
 4. NO FILLER: Start immediately. No introductions, no "Based on the data", no "I found this".
-5. GREETINGS: Respond to "hi/hello" with a single polite sentence.
-6. STRUCTURE: Use <b>HTML bolding</b> for key terms. Use bullet points ONLY if essential for clarity.
+5. FORMATTING: Use <b>HTML bolding</b> for all key terms, numbers, and deadlines.
+6. STRUCTURE: One line per topic using <br> for spacing. Avoid bulky paragraphs.
 
-TONE: Ultra-brief, literal, and factual."""
+EXAMPLE:
+<b>Maternity:</b> <b>90 days</b> paid (after <b>1yr</b> service). <b>30 days</b> stillbirth. Apply <b>5mo</b> early.
+
+TONE: Ultra-brief, literal, and surgical precision."""
 
 def _build_context(chunks: List[Dict[str, str]]) -> str:
     """Combines chunks into a raw string without technical headers."""
@@ -138,7 +139,7 @@ class LLMService:
         payload = {
             "model": target_model,
             "temperature": temp,
-            "max_tokens": 600,    # Further reduced for absolute TPM safety
+            "max_tokens": 1024,    # Increased to prevent cutoffs while maintaining TPM safety
             "messages": messages,
         }