app.py

import gradio as gr
from transformers import pipeline, AutoModelForCausalLM, AutoTokenizer
import torch
import theme

theme = theme.Theme()

import os
import sys
sys.path.append('../..')

#langchain
from langchain.text_splitter import RecursiveCharacterTextSplitter, CharacterTextSplitter
from langchain.embeddings import HuggingFaceEmbeddings
from langchain.prompts import PromptTemplate
from langchain.chains import RetrievalQA
from langchain.prompts import ChatPromptTemplate
from langchain.schema import StrOutputParser
from langchain.schema.runnable import Runnable
from langchain.schema.runnable.config import RunnableConfig
from langchain.chains import (
    LLMChain, ConversationalRetrievalChain)
from langchain.vectorstores import Chroma
from langchain.memory import ConversationBufferMemory
from langchain.chains import LLMChain
from langchain.prompts.prompt import PromptTemplate
from langchain.prompts.chat import ChatPromptTemplate, SystemMessagePromptTemplate
from langchain.prompts import SystemMessagePromptTemplate, HumanMessagePromptTemplate, ChatPromptTemplate,  MessagesPlaceholder
from langchain.document_loaders import PyPDFDirectoryLoader
from pydantic import BaseModel, Field
from langchain.output_parsers import PydanticOutputParser
from langchain_community.llms import HuggingFaceHub
from langchain_community.document_loaders import WebBaseLoader

import shutil



custom_title = "<span style='color: rgb(243, 239, 224);'>Green Greta</span>"


# Cell 1: Image Classification Model
image_pipeline = pipeline(task="image-classification", model="guillen/vit-basura-test1")

def predict_image(input_img):
    predictions = image_pipeline(input_img)
    return {p["label"]: p["score"] for p in predictions} 

image_gradio_app = gr.Interface(
    fn=predict_image,
    inputs=gr.Image(label="Image", sources=['upload', 'webcam'], type="pil"),
    outputs=[gr.Label(label="Result")],
    title=custom_title,
    theme=theme
)

loader = WebBaseLoader(["https://www.epa.gov/recycle/frequent-questions-recycling", "https://www.whitehorsedc.gov.uk/vale-of-white-horse-district-council/recycling-rubbish-and-waste/lets-get-real-about-recycling/", "https://www.teimas.com/blog/13-preguntas-y-respuestas-sobre-la-ley-de-residuos-07-2022", "https://www.molok.com/es/blog/gestion-de-residuos-solidos-urbanos-rsu-10-dudas-comunes"])
data=loader.load()
# split documents
text_splitter = RecursiveCharacterTextSplitter(
    chunk_size=1024,
    chunk_overlap=150,
    length_function=len
)
docs = text_splitter.split_documents(data)
# define embedding
embeddings = HuggingFaceEmbeddings(model_name='thenlper/gte-small')
# create vector database from data
persist_directory = 'docs/chroma/'

# Remove old database files if any
shutil.rmtree(persist_directory, ignore_errors=True)
vectordb = Chroma.from_documents(
    documents=docs,
    embedding=embeddings,
    persist_directory=persist_directory
)
# define retriever
retriever = vectordb.as_retriever(search_kwargs={"k": 2}, search_type="mmr")

class FinalAnswer(BaseModel):
    question: str = Field(description="the original question")
    answer: str = Field(description="the extracted answer")

# Assuming you have a parser for the FinalAnswer class
parser = PydanticOutputParser(pydantic_object=FinalAnswer)

template = """
Your name is AngryGreta and you are a recycling chatbot with the objective to anwer questions from user in English or Spanish /
Use the following pieces of context to answer the question /
If the question is English answer in English /
If the question is Spanish answer in Spanish /
Do not mention the word context when you answer a question /
Answer the question fully and provide as much relevant detail as possible. Do not cut your response short /
Context: {context}
User: {question}
{format_instructions}
"""

# Create the chat prompt templates
sys_prompt = SystemMessagePromptTemplate.from_template(template)
qa_prompt = ChatPromptTemplate(
    messages=[
        sys_prompt,
        HumanMessagePromptTemplate.from_template("{question}")],
    partial_variables={"format_instructions": parser.get_format_instructions()}
)
llm = HuggingFaceHub(
    repo_id="mistralai/Mixtral-8x7B-Instruct-v0.1",
    task="text-generation",
    model_kwargs={
        "max_new_tokens": 2000,
        "top_k": 30,
        "temperature": 0.1,
        "repetition_penalty": 1.03
    },
)

qa_chain = ConversationalRetrievalChain.from_llm(
    llm = llm,
    memory = ConversationBufferMemory(llm=llm, memory_key="chat_history", input_key='question', output_key='output'),
    retriever = retriever,
    verbose = True,
    combine_docs_chain_kwargs={'prompt': qa_prompt},
    get_chat_history = lambda h : h,
    rephrase_question = False,
    output_key = 'output',
)

def chat_interface(question,history):
    result = qa_chain.invoke({'question': question})
    output_string = result['output']

    # Find the index of the last occurrence of "answer": in the string
    answer_index = output_string.rfind('"answer":')

    # Extract the substring starting from the "answer": index
    answer_part = output_string[answer_index + len('"answer":'):].strip()

    # Find the next occurrence of a double quote to get the start of the answer value
    quote_index = answer_part.find('"')

    # Extract the answer value between double quotes
    answer_value = answer_part[quote_index + 1:answer_part.find('"', quote_index + 1)]
    
    return answer_value


chatbot_gradio_app = gr.ChatInterface(
    fn=chat_interface,
    title=custom_title
)

# Combine both interfaces into a single app
app = gr.TabbedInterface(
    [image_gradio_app, chatbot_gradio_app],
    tab_names=["Green Greta Image Classification","Green Greta Chat"],
    theme=theme
)

app.queue()
app.launch()