chain_setup.py

import os
from huggingface_hub import hf_hub_download
from langchain.llms import LlamaCpp
from langchain.chains import ConversationalRetrievalChain
from langchain.memory import ConversationBufferMemory

def load_llm():
    """
    Downloads the GGUF model for Arabic and loads it via llama-cpp.
    """
    model_file = hf_hub_download(
        repo_id="mobeidat/c4ai-command-r7b-arabic-02-2025-Q4_K_M-GGUF",
        filename="c4ai-command-r7b-arabic-02-2025-q4_k_m.gguf",
        local_dir="./models",
        local_dir_use_symlinks=False
    )

    # Pass an empty grammar file to bypass Jinja template parsing.
    llm = LlamaCpp(
        model_path=model_file,
        flash_attn=False,
        n_ctx=2048,
        n_batch=512,
        chat_format="chatml",
        grammar=None,
        streaming=True,
        grammar_path=None,  # ensure this file exists and is empty
        use_jinja=False,
        rope_freq_base=10000.0,
        rope_freq_scale=1.0,
        use_mmap=True,
        last_n_tokens_size=64,
        echo=False,
        repeat_penalty=1.1,
        temperature=0.8,
        top_k=40,
        top_p=0.95,
        logprobs=None,
        callback_manager=None,
        custom_get_token_ids = None,
        lora_base = None,
        lora_path = None,
        max_tokens = 256,
        metadata= None,
        n_gpu_layers= None,
        n_threads= None,
        stop=[],
        suffix= None,
        tags = None,
        use_mlock=False,
        vocab_only=False,
        logits_all= False,
        callbacks=None,
        f16_kv=True,
        n_parts=-1,
        seed=-1,
        verbose=True,
        client=None,
        cache=None
    )

    return llm

def build_conversational_chain(vectorstore):
    """
    Creates a ConversationalRetrievalChain using the local llama-cpp-based LLM 
    and a ConversationBufferMemory for multi-turn Q&A.
    """
    llm = load_llm()

    memory = ConversationBufferMemory(
        memory_key="chat_history",
        return_messages=True
    )

    qa_chain = ConversationalRetrievalChain.from_llm(
        llm=llm,
        retriever=vectorstore.as_retriever(search_type="similarity", search_kwargs={"k": 5}),
        memory=memory,
        verbose=True
    )

    return qa_chain