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, max_retries=3):
    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.
- If no general answer is possible, only then say: "No relevant information is available to answer this question."

🛑 DO NOT:
- Mention "context", "document", or "text" in any form.
- Use markdown, bullets, emojis, or code blocks.
- Use words like helpful, available, allowed, indemnified, excluded, etc.
- Use legal-style passive phrases like "shall be indemnified".
- Start with or include phrases like "Based on the context", "No details found in the context", etc.
- Give theory or long explanations.
- Say "This info is not covered", "Out of scope", or similar.
- Say "Please refer to other sources".
- Say "Not mentioned in the document" — if nothing is found, give a general insurance-based response.

✅ 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):
        for attempt in range(max_retries):
            try:
                genai.configure(api_key=key)
                model = genai.GenerativeModel("gemini-2.5-flash-lite")
                response = model.generate_content(prompt)
                response_text = getattr(response, "text", "").strip()

                if not response_text:
                    raise ValueError("Empty response received from Gemini API.")

                # 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()

                # Optional debug log:
                # print(f"[Gemini Response]: {response_text[:300]}")

                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
                    break  # Try next key
                else:
                    print(f"Retry {attempt+1}/{max_retries} failed using Gemini key {key[:8]}...: {e}")
                    last_exception = e
                    continue  # Retry same key

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