app.py

# -*- coding: utf-8 -*-
"""RAGLLAMA.ipynb

Automatically generated by Colab.

Original file is located at
    https://colab.research.google.com/drive/1z-I8OtD-uFaX9KHENLECbxMmPPuPDNSp
"""


from langchain.llms import HuggingFacePipeline
from transformers import AutoModelForCausalLM, AutoTokenizer, GenerationConfig, pipeline

model_name = "TheBloke/Llama-2-13b-Chat-GPTQ"

model = AutoModelForCausalLM.from_pretrained(model_name,
                                             device_map="auto",
                                             trust_remote_code=True)

tokenizer = AutoTokenizer.from_pretrained(model_name, use_fast=True)

gen_cfg = GenerationConfig.from_pretrained(model_name)
gen_cfg.max_new_tokens=512
gen_cfg.temperature=0.0000001 # 0.0
gen_cfg.return_full_text=True
gen_cfg.do_sample=True
gen_cfg.repetition_penalty=1.11

pipe=pipeline(
    task="text-generation",
    model=model,
    tokenizer=tokenizer,
    generation_config=gen_cfg
)

llm = HuggingFacePipeline(pipeline=pipe)

"""Testing LLM with Prompt Structure

"""

from textwrap import fill
from langchain.prompts import PromptTemplate

template = "Question: {text}"

prompt = PromptTemplate(
    input_variables=["text"],
    template=template,
)

text = "how a coffee bean is roasted?"
result = llm.invoke(prompt.format(text=text))
print(fill(result.strip(), width=100))

import locale
locale.getpreferredencoding = lambda: "UTF-8"

"""Load PDF"""

from langchain.document_loaders import UnstructuredPDFLoader
from langchain.vectorstores.utils import filter_complex_metadata  # 'filter_complex_metadata' removes complex metadata that are not in str, int, float or bool format

# Replace 'your_pdf_file.pdf' with the name of your uploaded PDF file
pdf_loader = UnstructuredPDFLoader('coffee.pdf')
pdf_doc = pdf_loader.load()
updated_pdf_doc = filter_complex_metadata(pdf_doc)

from langchain.text_splitter import RecursiveCharacterTextSplitter
text_splitter = RecursiveCharacterTextSplitter(chunk_size=1024, chunk_overlap=128)
chunked_pdf_doc = text_splitter.split_documents(updated_pdf_doc)
len(chunked_pdf_doc)

from langchain.embeddings import HuggingFaceEmbeddings
embeddings = HuggingFaceEmbeddings()

"""Creating Vector Database with FAISS (Similarity search)"""

# Commented out IPython magic to ensure Python compatibility.
# 
# %%time
# # Create the vectorized db with FAISS
# from langchain.vectorstores import FAISS
# db_pdf = FAISS.from_documents(chunked_pdf_doc, embeddings)
# 
# # Create the vectorized db with Chroma
# # from langchain.vectorstores import Chroma
# # db_pdf = Chroma.from_documents(chunked_pdf_doc, embeddings)

import transformers
transformers.logging.set_verbosity_error()

"""RetrivalQA - LLM Prompt Structure"""

# Commented out IPython magic to ensure Python compatibility.
# %%time
# from langchain.prompts import PromptTemplate
# from langchain.chains import RetrievalQA
# 
# # use the recommended propt style for the LLAMA 2 LLM
# prompt_template = """
# [INST] <>
# Use the following context to Answer the question at the end. Do not use any other information. If you can't find the relevant information in the context, just say you don't have enough information to answer the question. Don't try to make up an answer.
# 
# <>
# 
# {context}[/INST]
# 
# Question: {question}
# """
# 
# 
# prompt = PromptTemplate(template=prompt_template, input_variables=["context","question"])
# Chain_pdf = RetrievalQA.from_chain_type(
#     llm=llm,
#     chain_type="stuff",
#     # retriever=db.as_retriever(search_type="similarity_score_threshold", search_kwargs={'k': 5, 'score_threshold': 0.8})
#     # Similarity Search is the default way to retrieve documents relevant to a query, but we can use MMR by setting search_type = "mmr"
#     # k defines how many documents are returned; defaults to 4.
#     # score_threshold allows to set a minimum relevance for documents returned by the retriever, if we are using the "similarity_score_threshold" search type.
#     # return_source_documents=True, # Optional parameter, returns the source documents used to answer the question
#     retriever=db_pdf.as_retriever(), # (search_kwargs={'k': 5, 'score_threshold': 0.8}),
#     chain_type_kwargs={"prompt": prompt },
# )
# query = "what is a coffee?"
# result = Chain_pdf.invoke(query)
# result_text = result['result']
# answer_start_index = result_text.find("Question:")  # Find the start index of the answer
# answer = result_text[answer_start_index:].strip()  # Extract the answer
# print(fill(answer, width=100))
#

# Commented out IPython magic to ensure Python compatibility.
# %%time
# query = "explain about coffee brewing?"
# result = Chain_pdf.invoke(query)
# result_text = result['result']
# answer_start_index = result_text.find("Question:")
# answer = result_text[answer_start_index:].strip()
# print(fill(answer, width=100))

"""Hallucination Test

"""

# Commented out IPython magic to ensure Python compatibility.
# %%time
# query = "How to make a milk tea?"
# result = Chain_pdf.invoke(query)
# result_text = result['result']
# answer_start_index = result_text.find("Question:")
# answer = result_text[answer_start_index:].strip()
# print(fill(answer, width=100))



import gradio as gr

def query_qa(query):
    result = Chain_pdf.invoke(query)
    result_text = result['result']
    answer_start_index = result_text.find("Question:")
    answer = result_text[answer_start_index:].strip()
    return answer

qa_interface = gr.Interface(query_qa, "text", "text", title="Coffee Q&A")

qa_interface.launch()
debug=True