llm.py

import google.generativeai as genai
import os
import json
from dotenv import load_dotenv

load_dotenv()

# Support multiple Gemini keys (comma-separated or single key)
api_keys = os.getenv("GOOGLE_API_KEYS") or os.getenv("GOOGLE_API_KEY")
if not api_keys:
    raise ValueError("No Gemini API keys found in GOOGLE_API_KEYS or GOOGLE_API_KEY environment variable.")

api_keys = [k.strip() for k in api_keys.split(",") if k.strip()]
print(f"Loaded {len(api_keys)} Gemini API key(s)")

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 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.
- 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.
- 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.
- Dont Give Long theory Like Response Very Large Response Just Give Short And Good Response For The Question.
- If the question is general (math, code, tech, etc.) and No Matches With Context, answer normally without referencing the document.
- Avoid Saying  “Not found” or “Out of scope” For The Answer of The Question Try to Give Basic General Response For The Question.

🛑 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
- Over-explain or give long theory answers.
- Dont Give Directly In Answer Like "The provided information does not contain Details" or "The context does not provide information about this question." Instead, give a general answer if nothing is found in the context. If General Also Not Possible Then Give Like That (That Time Only Use That Response) "I am unable to provide an answer to this question based on the provided context. Please refer to the relevant policy documents or contact customer support for assistance."

✅ DO:
- Write in clean, informative language.
- Give complete answers in 2–3 sentences maximum.


📝 EXAMPLE ANSWERS:
- "Yes, the policy covers damage to personal property caused by fire, up to a limit of $50,000."
- "No, the policy does not cover pre-existing conditions."
- "The waiting period for coverage to begin is 30 days from the start date of the policy."


📤 OUTPUT FORMAT (strict):
Respond with only the following JSON — no explanations, no comments, no markdown:

{{
  "answers": [
    "Answer to question 1",
    "Answer to question 2",
    ...
  ]
}}

📚 CONTEXT:
{context}

❓ QUESTIONS:
{questions_text}

Your task: For each question, provide a complete, professional, and clearly written answer in 2–3 sentences using a formal but readable tone.
"""

    last_exception = None

    for key_idx, key in enumerate(api_keys):
        try:
            genai.configure(api_key=key)
            model = genai.GenerativeModel("gemini-2.5-flash-lite")
            response = model.generate_content(prompt)
            response_text = response.text.strip()

            # Clean JSON from wrapped code blocks
            if response_text.startswith("```json"):
                response_text = response_text.replace("```json", "").replace("```", "").strip()
            elif response_text.startswith("```"):
                response_text = response_text.replace("```", "").strip()

            parsed = json.loads(response_text)
            if "answers" in parsed and isinstance(parsed["answers"], list):
                return parsed
            else:
                raise ValueError("Invalid response format received from Gemini.")

        except Exception as e:
            msg = str(e).lower()
            if "429" in msg or "quota" in msg or "rate limit" in msg or "exceeded" in msg:
                print(f"Gemini key {key[:8]}... quota/rate limited. Trying next key ({key_idx+1}/{len(api_keys)}).")
                last_exception = e
                continue
            else:
                print(f"Error using Gemini key {key[:8]}...: {e}")
                return {"answers": [f"Error generating response: {str(e)}"] * len(questions)}

    print(f"All Gemini API keys failed. Last error: {last_exception}")
    return {"answers": [f"Error generating response: {str(last_exception)}"] * len(questions)}