vector_store_retriever.py

import gradio as gr
from langchain.document_loaders import DirectoryLoader, PyPDFLoader
from langchain.vectorstores import Chroma
from langchain.chains import RetrievalQA
from langchain.embeddings import HuggingFaceInstructEmbeddings
from langchain.agents import Tool
from langchain.text_splitter import RecursiveCharacterTextSplitter
from langchain.llms import HuggingFacePipeline
from transformers import LlamaTokenizer, LlamaForCausalLM, pipeline

# Load and process the text files
loader = DirectoryLoader('./new_papers/', glob="./*.pdf", loader_cls=PyPDFLoader)
documents = loader.load()

# Splitting the text into chunks
text_splitter = RecursiveCharacterTextSplitter(chunk_size=1000, chunk_overlap=200)
texts = text_splitter.split_documents(documents)

# HF Instructor Embeddings
instructor_embeddings = HuggingFaceInstructEmbeddings(model_name="hkunlp/instructor-xl", model_kwargs={"device": "cuda"})

# Embed and store the texts
persist_directory = 'db'
embedding = instructor_embeddings
vectordb = Chroma.from_documents(documents=texts, embedding=embedding, persist_directory=persist_directory)

# Make a retriever
retriever = vectordb.as_retriever(search_kwargs={"k": 3})

# Setup LLM for text generation
tokenizer = LlamaTokenizer.from_pretrained("TheBloke/wizardLM-7B-HF")
model = LlamaForCausalLM.from_pretrained("TheBloke/wizardLM-7B-HF", load_in_8bit=True, device_map='auto', torch_dtype=torch.float16, low_cpu_mem_usage=True)
pipe = pipeline("text-generation", model=model, tokenizer=tokenizer, max_length=1024, temperature=0, top_p=0.95, repetition_penalty=1.15)
local_llm = HuggingFacePipeline(pipeline=pipe)

# Make a chain
qa_chain = RetrievalQA.from_chain_type(llm=local_llm, chain_type="stuff", retriever=retriever, return_source_documents=True)

class VectorStoreRetrieverTool(Tool):
    name = "vectorstore_retriever"
    description = "This tool uses LangChain's RetrievalQA to find relevant answers from a vector store based on a given query."

    inputs = ["text"]
    outputs = ["text"]

    def __call__(self, query: str):
        # Run the query through the RetrievalQA chain
        llm_response = qa_chain(query)
        return llm_response['result']

# Create the Gradio interface using the HuggingFaceTool
tool = gr.Interface(
    VectorStoreRetrieverTool(),
    live=True,
    title="LangChain-Application: Vectorstore-Retriever",
    description="This tool uses LangChain's RetrievalQA to find relevant answers from a vector store based on a given query.",
)

# Launch the Gradio interface
tool.launch()