rag_pipeline.py

from pathlib import Path
from langchain.chains import RetrievalQA
from transformers import pipeline, AutoTokenizer
from langchain_community.vectorstores import Chroma
from langchain_community.document_loaders import DirectoryLoader
from langchain.text_splitter import RecursiveCharacterTextSplitter
from langchain_huggingface import HuggingFaceEmbeddings, HuggingFacePipeline

def load_rag_chain():
    pdf_dir = Path("data")
    loader = DirectoryLoader(str(pdf_dir), glob="*.pdf")
    text_splitter = RecursiveCharacterTextSplitter(chunk_size=1000, chunk_overlap=200)
    pages = loader.load_and_split(text_splitter=text_splitter)

    embeddings = HuggingFaceEmbeddings(
        model_name="sentence-transformers/all-MiniLM-L6-v2",
        model_kwargs={"device": "cpu"},
    )

    vectordb_dir = "chroma_db"
    vectordb = Chroma.from_documents(pages, embeddings, persist_directory=vectordb_dir)
    retriever = vectordb.as_retriever(search_type="mmr", search_kwargs={"k": 5})

    hf_pipeline = pipeline(
        "text-generation",
        model="mistralai/Mistral-7B-Instruct-v0.2",
        tokenizer=AutoTokenizer.from_pretrained("mistralai/Mistral-7B-Instruct-v0.2"),
        max_new_tokens=512,
        temperature=0.3,
        return_full_text=True,
        device=-1,
    )
    llm = HuggingFacePipeline(pipeline=hf_pipeline)

    qa_chain = RetrievalQA.from_llm(llm=llm, retriever=retriever)
    return qa_chain