llm_handler.py

import os
from dotenv import load_dotenv
from langchain_groq import ChatGroq

load_dotenv()

llm = ChatGroq(
    model_name="llama-3.3-70b-versatile",
    groq_api_key=os.getenv("GROQ_API_KEY"),
    temperature=0.3
)


def generate_answer(question, retrieved_chunks, language="arabic",
                     recent_history=None, conversation_summary=None):
    """
    Generates the final answer using Groq, grounded in retrieved
    document chunks and aware of conversation context.

    recent_history: list of the last 1-2 (question, answer) tuples,
                     kept in full detail
    conversation_summary: short summary of everything OLDER than
                     the recent history, kept compact
    """

    context = "\n\n".join(retrieved_chunks)

    history_text = ""
    if conversation_summary:
        history_text += f"Summary of earlier conversation:\n{conversation_summary}\n\n"
    if recent_history:
        history_text += "Most recent exchange(s):\n"
        for past_question, past_answer in recent_history:
            history_text += f"Q: {past_question}\nA: {past_answer}\n\n"

    if language == "arabic":
        prompt = f"""أنت مساعد ذكي متخصص في الإجابة على الأسئلة بناءً على النص المعطى فقط.

{history_text if history_text else ""}

السياق المتاح من الوثيقة:
{context}

السؤال الحالي:
{question}

التعليمات:
- أجب باللغة العربية فقط
- إذا كان السؤال يشير إلى محادثة سابقة (مثل "ذلك" أو "هذا")، استخدم ملخص المحادثة أو آخر تبادل لفهم المرجع
- استخدم فقط المعلومات الموجودة في السياق أعلاه
- لا تضف أي معلومات غير موجودة في الوثيقة، حتى لو كانت معلومات عامة صحيحة
- إذا لم تجد الإجابة، قل بوضوح أنك لا تملك معلومات كافية

الإجابة:"""
    else:
        prompt = f"""You are a helpful assistant that answers questions based on the provided document context.

{history_text if history_text else ""}

Document context (in Arabic):
{context}

Current question:
{question}

Instructions:
- Answer ONLY in English
- If the question refers to earlier conversation (like "that" or "it"), use the summary or recent exchange to understand the reference
- Use the information found in the context above
- If the answer is not in the context, clearly say you don't have enough information
- Don't give information which is not in the PDF if something is not there in the PDF simply say you don't have enough information

Answer:"""

    response = llm.invoke(prompt)
    return response.content


def translate_to_arabic(english_text):
    """
    Translates an English question into Arabic using Groq,
    so we can search our all-Arabic vector store accurately.
    This fixes cross-lingual retrieval failures on larger,
    more complex documents.
    """
    prompt = f"""Translate the following English question into Modern Standard Arabic.
Only output the Arabic translation, nothing else - no explanation, no quotes.

English: {english_text}

Arabic translation:"""

    response = llm.invoke(prompt)
    return response.content.strip()


def summarize_conversation(existing_summary, old_question, old_answer):
    """
    Takes the current running summary plus one older exchange
    that's about to fall out of the "recent" window, and asks
    Groq to fold it into an updated, still-short summary.

    This keeps conversation memory compact regardless of how
    long the chat gets, instead of sending the full raw history
    every time.
    """
    prompt = f"""You are maintaining a brief running summary of a conversation between a user and an assistant about a document.

Existing summary so far:
{existing_summary if existing_summary else "(no summary yet - this is the first exchange to summarize)"}

New exchange to fold in:
User asked: {old_question}
Assistant answered: {old_answer}

Update the summary to include this new exchange, staying concise.
Keep it to 2-4 sentences maximum. Focus on WHAT TOPICS were discussed
and any specific facts/numbers that might be referenced later, not
the exact wording. Output ONLY the updated summary, nothing else.

Updated summary:"""

    response = llm.invoke(prompt)
    return response.content.strip()