confluence_rag.py

import os
from typing import List

from langchain_community.document_loaders import UnstructuredPDFLoader
from langchain_community.embeddings import HuggingFaceInferenceAPIEmbeddings
from langchain_community.llms.huggingface_endpoint import HuggingFaceEndpoint
from langchain.prompts import ChatPromptTemplate
from langchain.schema.output_parser import StrOutputParser
from langchain.schema.runnable import RunnablePassthrough
from langchain.text_splitter import RecursiveCharacterTextSplitter
from langchain_community.vectorstores.chroma import Chroma
from langchain_core.runnables.base import RunnableSequence
from langchain_core.vectorstores import VectorStoreRetriever

from dotenv import load_dotenv


load_dotenv()
HF_API_KEY = os.environ["HF_API_KEY"]


class MistralOutputParser(StrOutputParser):
    """OutputParser that parser llm result from Mistral API"""

    def parse(self, text: str) -> str:
        """
        Returns the input text with no changes.

        Args:
            text (str): text to parse

        Returns:
            str: parsed text
        """
        return text.split("[/INST]")[-1].strip()


def load_pdf(
    document_path: str,
    mode: str = "single",
    strategy: str = "fast",
    chunk_size: int = 500,
    chunk_overlap: int = 0,
) -> List[str]:
    """
    Load a pdf document and split it into chunks of text.

    Args:
        document_path (Path): path to the pdf document
        mode (str, optional): mode of the loader. Defaults to "single".
        strategy (str, optional): strategy of the loader. Defaults to "fast".
        chunk_size (int, optional): size of the chunks. Defaults to 500.
        chunk_overlap (int, optional): overlap of the chunks. Defaults to 0.

    Returns:
        List[str]: list of chunks of text
    """

    # Load the document
    loader = UnstructuredPDFLoader(
        document_path,
        mode=mode,
        strategy=strategy,
    )

    docs = loader.load()

    # Split the document into chunks of text
    text_splitter = RecursiveCharacterTextSplitter(
        chunk_size=chunk_size, chunk_overlap=chunk_overlap
    )
    all_splits = text_splitter.split_documents(docs)

    return all_splits


def store_vector(all_splits: List[str]) -> VectorStoreRetriever:
    """
    Store vector of each chunk of text.

    Args:
        all_splits (List[str]): list of chunks of text

    Returns:
        VectorStoreRetriever: retriever that can be used to retrieve the vector of a chunk of text
    """

    # Use the HuggingFace distilbert-base-uncased model to embed the text
    embeddings_model_url = (
        # "https://api-inference.huggingface.co/models/distilbert-base-uncased"
        "https://api-inference.huggingface.co/models/Salesforce/SFR-Embedding-Mistral"
    )

    embeddings = HuggingFaceInferenceAPIEmbeddings(
        endpoint_url=embeddings_model_url,
        api_key=HF_API_KEY,
    )

    # Store the embeddings of each chunk of text into ChromaDB
    vector_store = Chroma.from_documents(all_splits, embeddings)
    retriever = vector_store.as_retriever()

    return retriever


def generate_mistral_rag_prompt() -> ChatPromptTemplate:
    """
    Generate a prompt for Mistral API wiht RAG.

    Returns:
        ChatPromptTemplate: prompt for Mistral API
    """
    template = "<s>[INST] {context} {prompt} [/INST]"
    prompt_template = ChatPromptTemplate.from_template(template)
    return prompt_template


def generate_mistral_simple_prompt() -> ChatPromptTemplate:
    """
    Generate a simple prompt for Mistral without RAG.

    Returns:
        ChatPromptTemplate: prompt for Mistral API
    """
    template = "[INST] {prompt} [/INST]"
    prompt_template = ChatPromptTemplate.from_template(template)
    return prompt_template


def generate_rag_chain(retriever: VectorStoreRetriever = None) -> RunnableSequence:
    """
    Generate a RAG chain with Mistral API and ChromaDB.

    Args:
        Retriever (VectorStoreRetriever): retriever that can be used to retrieve the vector of a chunk of text

    Returns:
        RunnableSequence: RAG chain
    """
    # Use the Mistral Free prototype API
    mistral_url = (
        # "https://api-inference.huggingface.co/models/mistralai/Mistral-7B-Instruct-v0.2"
        "https://api-inference.huggingface.co/models/mistralai/Mixtral-8x22B-Instruct-v0.1"
    )

    model_endpoint = HuggingFaceEndpoint(
        endpoint_url=mistral_url,
        huggingfacehub_api_token=HF_API_KEY,
        task="text2text-generation",
        max_new_tokens=1024
    )

    # Use a custom output parser
    output_parser = MistralOutputParser()

    # If no retriever is provided, use a simple prompt
    if retriever is None:
        entry = {"prompt": RunnablePassthrough()}
        return entry | generate_mistral_simple_prompt() | model_endpoint | output_parser

    # If a retriever is provided, use a RAG prompt
    retrieval = {"context": retriever, "prompt": RunnablePassthrough()}

    return retrieval | generate_mistral_rag_prompt() | model_endpoint | output_parser


def load_multiple_pdf(document_paths: List[str]) -> List[str]:
    """
    Load multiple pdf documents and split them into chunks of text.

    Args:
        document_paths (List[str]): list of paths to the pdf documents

    Returns:
        List[str]: list of chunks of text
    """
    docs = []
    for document_path in document_paths:
        loader = UnstructuredPDFLoader(
            document_path,
            mode="single",
            strategy="fast",
        )
        docs.extend(loader.load())

    text_splitter = RecursiveCharacterTextSplitter(chunk_size=500, chunk_overlap=25)
    all_splits = text_splitter.split_documents(docs)
    return all_splits