PDFChat / app.py
AkashDataScience's picture
Interface update
078bb1a
raw
history blame
5.77 kB
import os
import gradio as gr
from pinecone import Pinecone
from langchain_chroma import Chroma
from langchain_core.prompts import PromptTemplate
from langchain_pinecone import PineconeVectorStore
from langchain_community.vectorstores import FAISS
from langchain_community.vectorstores import LanceDB
from google.api_core.exceptions import ResourceExhausted
from langchain_text_splitters import CharacterTextSplitter
from langchain_community.document_loaders import PyPDFLoader
from langchain_google_genai import GoogleGenerativeAIEmbeddings, GoogleGenerativeAI
embeddings = GoogleGenerativeAIEmbeddings(model="models/text-embedding-004")
gemini = GoogleGenerativeAI(model="models/gemini-2.0-flash", temperature=0.0, top_k=1, top_p=0.0)
prompt_template = """
Context:\n {context}?\n
Question: \n{question}\n
Answer:
"""
prompt = PromptTemplate(template = prompt_template, input_variables = ["context", "question"])
chain = prompt | gemini.with_retry(retry_if_exception_type=(ResourceExhausted,))
index_name = "langchain-test-index"
def extract_text_from_pdf(pdf_path):
raw_documents = []
for path in pdf_path:
raw_documents.extend(PyPDFLoader(path).load())
return raw_documents
def chunk_text(raw_documents, chunk_size, chunk_overlap):
text_splitter = CharacterTextSplitter(chunk_size=chunk_size, chunk_overlap=chunk_overlap)
documents = text_splitter.split_documents(raw_documents)
return documents
def delete_pinecone():
pc = Pinecone(api_key=os.environ["PINECONE_API_KEY"])
index = pc.Index(host="https://langchain-test-index-la2n80y.svc.aped-4627-b74a.pinecone.io")
if index.describe_index_stats()['total_vector_count'] > 0:
index.delete(delete_all=True)
def store_chroma_db(documents):
chroma_db = Chroma.from_documents(documents, embeddings, persist_directory="./chroma_db")
def store_faiss_db(documents):
faiss_db = FAISS.from_documents(documents, embeddings)
faiss_db.save_local("./faiss_db")
def store_lance_db(documents):
lance_db = LanceDB.from_documents(documents, embeddings, uri="./lance_db")
def store_pinecone_db(documents):
pinecone_db = PineconeVectorStore.from_documents(documents, index_name=index_name,
embedding=embeddings)
def load_chroma_db():
chroma_db = Chroma(persist_directory="./chroma_db", embedding_function=embeddings)
return chroma_db
def load_faiss_db():
faiss_db = FAISS.load_local("./faiss_db", embeddings, allow_dangerous_deserialization=True)
return faiss_db
def load_lance_db():
lance_db = LanceDB(embedding=embeddings, uri="./lance_db")
return lance_db
def connect_pinecone_db():
pinecone_db = PineconeVectorStore(index_name=index_name, embedding=embeddings)
return pinecone_db
def invoke_chain(db, query):
docs = db.similarity_search(query)
answer = chain.invoke({"context":docs, "question": query}, return_only_outputs=True)
return answer
def store_embeddings(pdf_path, chunk_size, chunk_overlap):
raw_documents = extract_text_from_pdf(pdf_path)
documents = chunk_text(raw_documents, chunk_size, chunk_overlap)
delete_pinecone()
store_chroma_db(documents)
store_faiss_db(documents)
store_lance_db(documents)
store_pinecone_db(documents)
return "All embeddings are stored in vector database"
title = "PDF Chat"
description = "A simple Gradio interface to query PDFs and compare vector database"
examples = [[["data/amazon-10-k-2024.pdf"], 1000, 100],
[["data/goog-10-k-2023.pdf"], 1000, 100]]
def inference(query):
chroma_db = load_chroma_db()
chroma_answer = invoke_chain(chroma_db, query)
faiss_db = load_faiss_db()
faiss_answer = invoke_chain(faiss_db, query)
lance_db = load_lance_db()
lance_answer = invoke_chain(lance_db, query)
pinecone_db = connect_pinecone_db()
pinecoce_answer = invoke_chain(pinecone_db, query)
return chroma_answer, faiss_answer, lance_answer, pinecoce_answer
with gr.Blocks(theme=gr.themes.Monochrome()) as demo:
gr.Markdown(f"# {title}\n{description}")
with gr.Row():
with gr.Column():
pdf = gr.File(label="Input PDFs", file_count="multiple", file_types=[".pdf"])
chunk_size = gr.Slider(0, 2000, 1000, 100, label="Size of Chunk")
chunk_overlap = gr.Slider(0, 1000, 100, 100, label="Size of Chunk Overlap")
with gr.Row():
clear_btn = gr.ClearButton(components=[pdf, chunk_size, chunk_overlap])
submit_btn = gr.Button("Store Embeddings", variant='primary')
with gr.Column():
message = gr.Textbox(label="Status", type="text")
submit_btn.click(store_embeddings, inputs=[pdf, chunk_size, chunk_overlap], outputs=message)
with gr.Row():
with gr.Column():
chroma_out = gr.Textbox(label="ChromaDB Response", type="text")
faiss_out = gr.Textbox(label="FaissDB Response", type="text")
with gr.Column():
lance_out = gr.Textbox(label="LanceDB Response", type="text")
pinecone_out = gr.Textbox(label="PineconeDB Response", type="text")
with gr.Row():
with gr.Column():
text = gr.Textbox(label="Question", type="text")
with gr.Row():
chat_clear_btn = gr.ClearButton(components=[text])
chat_submit_btn = gr.Button("Submit", variant='primary')
chat_submit_btn.click(inference, inputs=[text], outputs=[chroma_out, faiss_out, lance_out,
pinecone_out])
examples_obj = gr.Examples(examples=examples, inputs=[pdf, chunk_size, chunk_overlap])
demo.launch()