app.py

import gradio as gr
from langchain import HuggingFaceHub
from langchain.chains.question_answering import load_qa_chain
from langchain.document_loaders import PyMuPDFLoader
from langchain.embeddings import HuggingFaceEmbeddings
from langchain.llms import HuggingFacePipeline
from langchain.text_splitter import RecursiveCharacterTextSplitter
from langchain.vectorstores import FAISS


# Number of search results to query from the vector database.
SIMILARITY_SEARCH_COUNT = 2

# Size of each document chunk in number of characters.
CHUNK_SIZE = 1000

# Maximum number of output tokens.
MODEL_MAX_LENGTH = 300


print("Loading documents")
loader = PyMuPDFLoader("rdna3-shader-instruction-set-architecture-feb-2023_0.pdf")
documents = loader.load()

print("Creating chunks")
splitter = RecursiveCharacterTextSplitter(chunk_size=CHUNK_SIZE, chunk_overlap=0)
chunks = splitter.split_documents(documents)

print("Creating database")
embeddings = HuggingFaceEmbeddings()
db = FAISS.from_documents(chunks, embeddings)

print("Loading model")
llm = HuggingFacePipeline.from_model_id(
    model_id="google/flan-t5-large",
    task="text2text-generation",
    model_kwargs={"temperature": 0, "max_length": MODEL_MAX_LENGTH})
chain = load_qa_chain(llm, chain_type="stuff")

def ask(question):
    answers = db.similarity_search(question, k=SIMILARITY_SEARCH_COUNT)
    result = chain.run(input_documents=answers, question=question)
    return result

# Warm up.
ask("What is VGPR")

iface = gr.Interface(
    fn=ask,
    inputs=gr.Textbox(label="Question", placeholder="What is..."),
    outputs=gr.Textbox(label="Answer"),
    allow_flagging=False)

iface.launch(share=False)