[Edit] Udpate Prompt New

.gitignore

@@ -25,7 +25,7 @@ wheels/
 .installed.cfg
 *.egg
 MANIFEST
-
+.cache
 # Virtual environments
 venv/
 env/

llm.py

@@ -16,29 +16,30 @@ def query_gemini(questions, contexts):
         context = "\n\n".join(contexts)
         questions_text = "\n".join([f"{i+1}. {q}" for i, q in enumerate(questions)])
         prompt = f"""
-You are an expert insurance assistant helping users understand policy details. Use the information provided to answer each question clearly and completely — as if you're speaking directly to the user. Do not reference the document or say where the information came from.
+You are an expert insurance assistant generating formal yet user-facing answers to policy questions. Your goal is to write professional, structured answers that reflect the language of policy documents — but are still human-readable and easy to understand.
 
-🎯 GOAL:
-- Each answer must be fully informative and written in natural, easy-to-understand language.
-- Aim for 2–3 clear and complete sentences per answer.
-- Rephrase complex clauses into simpler terms. If conditions, eligibility, waiting periods, or limits apply — clearly mention them.
-- If the answer isn’t directly found, respond using general insurance knowledge in a confident and helpful tone.
+🧠 FORMAT & TONE GUIDELINES:
+- Write in **professional third-person language** (no "you", no "we").
+- Use **clear sentence structure** with proper punctuation and spacing.
+- Do NOT write in legalese or robotic passive constructions.
+- Include eligibility, limits, and waiting periods explicitly where relevant.
+- Keep it factual, neutral, and easy to follow.
+- Dont be Too Long or  Repeats Unnecessary Information In the response.
+- Answer Question Based On The Context and If Not Context Matches Then Answer Based On General Knowledge.
 
-📘 INSTRUCTIONS:
-1. DO NOT mention the words “context,” “provided text,” “document,” or similar. Just answer directly.
-2. Avoid robotic or legalistic tone. Use human, natural phrasing.
-3. Do NOT include legal codes, section numbers, or abbreviations like “as per...” or “clause 3.2.1”.
-4. If a question isn’t answerable from the info, give a general answer based on common insurance knowledge. Be confident and clear.
-5. Never leave the answer empty — say something helpful or general if the exact info isn't available.
-6. Follow this tone and format style:
+🛑 DO NOT:
+- Use words like "context", "document", or "text".
+- Output markdown, bullets, emojis, or markdown code blocks.
+- Say "helpful", "available", "allowed", "indemnified", "excluded", etc.
+- Use overly robotic passive constructions like "shall be indemnified".
 
-✅ GOOD EXAMPLES:
-- "A grace period of thirty days is allowed after the due date for premium payment, so the policy can be renewed without losing benefits."
-- "Yes, maternity expenses are covered after 24 months of continuous coverage, and are limited to two deliveries or terminations during the policy period."
-- "Donor hospitalization costs are covered when the organ is used for the insured person, provided the transplant is legally compliant."
+✅ DO:
+- Write in clean, informative language.
+- Give complete answers in 2–3 sentences maximum.
+- Provide fallback answers if exact info is not available.
 
-📤 RETURN FORMAT:
-Respond in the exact JSON format below — no extra comments or explanation.
+📤 OUTPUT FORMAT (strict):
+Respond with **only** the following JSON — no explanations, no comments, no markdown:
 
 {{
   "answers": [
@@ -54,10 +55,11 @@ Respond in the exact JSON format below — no extra comments or explanation.
 ❓ QUESTIONS:
 {questions_text}
 
-Your task: For each question, provide a clear, helpful answer in 2–3 sentences. Use natural, professional language — never mention the context or document, just explain as if you know the answer.
+Your task: For each question, provide a complete, professional, and clearly written answer in 2–3 sentences using a formal but readable tone.
 """
 
 
+        
 
         model = genai.GenerativeModel('gemini-2.5-flash-lite')
         response = model.generate_content(prompt)