rag_engine.py

from langchain_groq import ChatGroq
from langchain_core.prompts import ChatPromptTemplate
from langchain_core.output_parsers import StrOutputParser
from langchain_core.runnables import RunnablePassthrough
import config

def initialize_llm():
    """
    Initialize Groq LLM.
    """
    return ChatGroq(
        model=config.LLM_MODEL_NAME,
        temperature=0.1,
        max_tokens=2000,
        api_key=config.GROQ_API_KEY
    )

def get_rag_prompt():
    """
    Create the prompt template for RAG.
    """
    return ChatPromptTemplate.from_template("""
You are a Senior Legal Consultant specializing in the laws of Pakistan.

CONTEXT:
1. Constitution of Pakistan
2. Pakistan Penal Code

INSTRUCTIONS:
- Adoption a formal, professional, and authoritative tone suitable for legal memoranda.
- Cite specific Articles, Sections, or Clauses extensively.
- If the information is present: Provide a direct, concise legal opinion.
- If the information is MISSING: State clearly "The provided legal documents do not contain specific provisions regarding [topic]." Do not apologize.
- Structure your response with clear headings if necessary.

LEGAL CONTEXT:
{context}

QUERY: {question}

LEGAL OPINION:
""")

def format_docs(docs):
    """
    Format retrieved documents for the prompt.
    """
    formatted = []
    for i, doc in enumerate(docs):
        source = doc.metadata.get('source', 'Unknown Document')
        page = doc.metadata.get('page', 'N/A')
        # Limit content length to avoid context window issues, though Groq usually has large context
        content = doc.page_content[:800] 
        formatted.append(f"[Document {i+1}: {source}, Page {page}]")
        formatted.append(content)
        formatted.append("-" * 50)
    return "\n".join(formatted)

def create_rag_chain(retriever):
    """
    Build the primary RAG chain.
    """
    llm = initialize_llm()
    prompt = get_rag_prompt()
    
    chain = (
        {"context": retriever | format_docs, "question": RunnablePassthrough()}
        | prompt
        | llm
        | StrOutputParser()
    )
    return chain