Add Good Prompt

embedder.py

@@ -3,22 +3,19 @@ import numpy as np
 import os
 from sentence_transformers import SentenceTransformer
 
-# Use a local cache for transformer downloads
 cache_dir = os.path.join(os.getcwd(), ".cache")
 os.makedirs(cache_dir, exist_ok=True)
 os.environ['HF_HOME'] = cache_dir
 os.environ['TRANSFORMERS_CACHE'] = cache_dir
 
-# Lazy-loaded model
 _model = None
 
-def preload_model(model_name="all-MiniLM-L6-v2"):
+def preload_model(model_name="paraphrase-MiniLM-L3-v2"):
     global _model
     if _model is not None:
         return _model
 
-    print("Preloading sentence transformer model...")
-
+    print(f"Preloading sentence transformer model: {model_name}...")
     try:
         _model = SentenceTransformer(model_name, cache_folder=cache_dir)
     except Exception as e:
@@ -35,18 +32,14 @@ def get_model():
 
 def build_faiss_index(chunks, batch_size=128, show_progress_bar=False):
     model = get_model()
-
-    # Encode using batching for speed
     embeddings = model.encode(
         chunks,
         batch_size=batch_size,
         show_progress_bar=show_progress_bar,
         convert_to_numpy=True,
-        normalize_embeddings=True  # Helps FAISS L2 perform better
+        normalize_embeddings=True
     )
-
     dim = embeddings.shape[1]
     index = faiss.IndexFlatL2(dim)
     index.add(embeddings)
-
     return index, chunks

llm.py

@@ -2,6 +2,7 @@ import google.generativeai as genai
 import os
 import json
 from dotenv import load_dotenv
+
 load_dotenv()
 
 api_key = os.getenv("GOOGLE_API_KEY")
@@ -16,30 +17,36 @@ 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 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.
+You are an expert insurance assistant generating formal yet user-facing answers to policy questions and Other Human 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.
 
 🧠 FORMAT & TONE GUIDELINES:
-- Write in **professional third-person language** (no "you", no "we").
-- Use **clear sentence structure** with proper punctuation and spacing.
+- 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.
+- First, try to answer each question using information from the provided context.
+- If the question is NOT covered by the context Provide Then Give The General Answer It Not Be In Context if Nothing Found Give Normal Ai Answer for The Question Correctly
+- Limit each answer to 2–3 sentences, and do not repeat unnecessary information.
+- If a question can be answered with a simple "Yes", "No", "Can apply", or "Cannot apply", then begin the answer with that phrase, followed by a short supporting Statement In Natural Human Like response.So Give A Good Answer For The Question With Correct Information.
+- Avoid giving  theory Based Long Long answers Try to Give Short Good Reasonable Answers.
 
 🛑 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".
+- Dont Give In Message Like "Based On The Context "Or "Nothing Refered In The context" Like That Dont Give In Response Try To Give Answer For The Question Alone
 
 ✅ DO:
 - Write in clean, informative language.
 - Give complete answers in 2–3 sentences maximum.
-- Provide fallback answers if exact info is not available.
+
+
+
 
 📤 OUTPUT FORMAT (strict):
-Respond with **only** the following JSON — no explanations, no comments, no markdown:
+Respond with only the following JSON — no explanations, no comments, no markdown:
 
 {{
   "answers": [
@@ -57,10 +64,6 @@ Respond with **only** the following JSON — no explanations, no comments, no ma
 
 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)
         response_text = response.text.strip()

retriever.py

@@ -1,12 +1,11 @@
-from sentence_transformers import SentenceTransformer
-import numpy as np
-import time
+from sentence_transformers.util import cos_sim
 from embedder import get_model
+import numpy as np
+
+def retrieve_chunks(index, texts, question, top_k=15):
+    model = get_model()
+    q_embedding = model.encode([question], convert_to_numpy=True, normalize_embeddings=True)[0]
 
-# Use the preloaded model from embedder instead of creating a new instance
-def retrieve_chunks(index, texts, query, k=15):
-    model = get_model()  # Use the preloaded model
-    query_vec = model.encode([query])
-    distances, indices = index.search(np.array(query_vec), k)
-    results = [texts[i] for i in indices[0]]
-    return results
+    scores, indices = index.search(np.array([q_embedding]), top_k)
+    selected = [texts[i] for i in indices[0]]
+    return selected