src/simple_rag.py

# Modified RAG Pipeline for General Document Q&A (Khmer & English)

import os
import logging
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM, pipeline, AutoModel
from langchain.text_splitter import RecursiveCharacterTextSplitter
from langchain.schema import Document
from langchain.vectorstores.chroma import Chroma
from langchain.embeddings import HuggingFaceEmbeddings
from langchain.document_loaders import PyPDFDirectoryLoader
from openai import OpenAI

logging.basicConfig(level=logging.INFO)

use_gpu = torch.cuda.is_available()

if use_gpu:
    print("CUDA device in use:", torch.cuda.get_device_name(0))
else:
    print("Running on CPU. No GPU detected.")

# Load API key from HF Space secrets
SEALION_API_KEY = os.environ.get("SEALION_API_KEY")

client = OpenAI(
    api_key=SEALION_API_KEY,
    base_url="https://api.sea-lion.ai/v1"
)

# Use Hugging Face's writable directory
WRITABLE_DIR = os.environ.get("HOME", "/app")

DATA_PATH = os.path.join(WRITABLE_DIR, "src", "data")
CHROMA_PATH = os.path.join(WRITABLE_DIR, "src", "chroma")
embedding_model = HuggingFaceEmbeddings(model_name="intfloat/multilingual-e5-base")

# Generic assistant prompt for dual Khmer/English
PROMPT_TEMPLATE = """
Respond ONLY in the same language as the question. 
If the question is in English, answer in English. 
If the question is in Khmer, answer in Khmer.

You are a helpful assistant. 
Use only the provided context below to answer the question. 
Do not mention the context or that you used it.

Context:
{context}

Question:
{question}

Answer:
"""


def load_documents():
    loader = PyPDFDirectoryLoader(DATA_PATH)
    return loader.load()

def split_text(documents: list[Document]):
    splitter = RecursiveCharacterTextSplitter(
        chunk_size=512, chunk_overlap=100, length_function=len, add_start_index=True
    )
    chunks = splitter.split_documents(documents)
    logging.info(f"Split {len(documents)} documents into {len(chunks)} chunks.")
    return chunks

def save_to_chroma(chunks: list[Document]):
    if os.path.exists(CHROMA_PATH):
        db = Chroma(persist_directory=CHROMA_PATH, embedding_function=embedding_model)
        db.add_documents(chunks)
        logging.info("Added documents to existing Chroma DB.")
    else:
        db = Chroma.from_documents(
            chunks, embedding_model, persist_directory=CHROMA_PATH
        )
        logging.info("Created new Chroma DB.")
    db.persist()
    logging.info(f"Saved {len(chunks)} chunks to Chroma.")

def generate_data_store():
    documents = load_documents()
    chunks = split_text(documents)
    save_to_chroma(chunks)

def ask_question(query_text: str, k: int = 3):
    logging.info("Processing user question...")

    db = Chroma(persist_directory=CHROMA_PATH, embedding_function=embedding_model)
    results = db.similarity_search(query_text, k=k)

    context_chunks = []
    for doc in results:
        meta = doc.metadata or {}
        context_chunks.append({
            "filename": os.path.basename(meta.get("source", "unknown.pdf")),
            "page": meta.get("page", 1),
            "text": doc.page_content.strip()
        })

    context_text = "\n\n".join(chunk["text"] for chunk in context_chunks)
    prompt = PROMPT_TEMPLATE.format(context=context_text, question=query_text)

    messages = [{"role": "user", "content": prompt}]
    logging.info("Sending prompt to model...")
    
    try:
        logging.info("Sending prompt to SEA-LION API...")
        completion = client.chat.completions.create(
            model="aisingapore/Llama-SEA-LION-v3.5-8B-R",
            messages=messages,
            extra_body={
                "chat_template_kwargs": {
                    "thinking_mode": "off"
                },
                "cache": {
                    "no-cache": True
                }
            },
            
            max_tokens=512
        )
        answer = completion.choices[0].message.content.strip()
    except Exception as e:
        logging.error(f"Error calling SEA-LION API: {e}")
        answer = "Sorry, something went wrong when contacting the language model."

    return answer, context_chunks