app/core/llm_engine.py

# llm_engine.py

import google.generativeai as genai
from app.core.config import GEMINI_API_KEY
from langchain_core.prompts import PromptTemplate
from langchain_core.output_parsers import StrOutputParser
from langchain_google_genai import ChatGoogleGenerativeAI

# ✅ Configure Gemini client
genai.configure(api_key=GEMINI_API_KEY)


def ask_gemini(context: str, question: str) -> str:
    """
    Ask Gemini a question based on document context using LangChain for better formatting and control.
    """

    try:
        # ✅ Initialize Gemini LLM via LangChain
        llm = ChatGoogleGenerativeAI(
            model="gemini-2.5-flash",
            google_api_key=GEMINI_API_KEY,
            temperature=0.4,
            max_output_tokens=2048,
            convert_system_message_to_human=True
        )

        # ✅ Define a structured, formatting-rich prompt
        prompt = PromptTemplate(
            input_variables=["context", "question"],
            template=(
                "You are an intelligent document assistant.\n"
                "Answer the user's question strictly using the provided context.\n"
                "Respond in **clean Markdown formatting** with:\n"
                "- Headings (##)\n"
                "- Bullet points and numbered lists\n"
                "- **Bold keywords**\n"
                "- Tables (if useful)\n"
                "- Code blocks when necessary\n"
                "- Proper spacing and paragraphs for readability\n\n"
                "### 📄 Document Context:\n{context}\n\n"
                "### 💬 User Question:\n{question}\n\n"
                "### 🧠 Answer:"
            )
        )

        # ✅ Combine the prompt, model, and parser (modern LCEL chain)
        chain = prompt | llm | StrOutputParser()

        # ✅ Run the chain
        response = chain.invoke({"context": context, "question": question})

        return response.strip() if response else "⚠️ No response from Gemini."

    except Exception as e:
        return f"⚠️ Gemini (LangChain) error: {str(e)}"