File size: 4,208 Bytes
e4d6763 6d5de73 e4d6763 2b4effb e4d6763 2b4effb e4d6763 6d5de73 d4b5c04 e4d6763 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 |
import os
import streamlit as st
import pickle
import time
from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig
from langchain.llms.base import LLM
from langchain.chains import RetrievalQAWithSourcesChain
from langchain.text_splitter import RecursiveCharacterTextSplitter
from langchain.document_loaders import UnstructuredURLLoader
from langchain.vectorstores import FAISS
from langchain.embeddings import HuggingFaceEmbeddings
from huggingface_hub import login
# Login to Hugging Face
login(os.getenv('HF_llama3chat8b'))
class CustomHuggingFaceLLM(LLM):
def __init__(self, model_name, temperature=0.7):
# Configure 8-bit quantization using `BitsAndBytesConfig`
quantization_config = BitsAndBytesConfig(
load_in_8bit=True, # Enable 8-bit quantization
llm_int8_enable_fp32_cpu_offload=True # Offload FP32 operations to CPU for further memory savings
)
self.model = AutoModelForCausalLM.from_pretrained(model_name, device_map="cpu", quantization_config=quantization_config)
self.tokenizer = AutoTokenizer.from_pretrained(model_name)
self.temperature = temperature
def _call(self, prompt, stop=None):
input_ids = self.tokenizer.encode(prompt, return_tensors="pt")
output = self.model.generate(
input_ids,
max_length=512,
temperature=self.temperature,
do_sample=True,
top_p=0.95,
top_k=3
)
generated_text = self.tokenizer.decode(output[0], skip_special_tokens=True)
return generated_text
@property
def _identifying_params(self):
return {"model_name": self.model.config._name_or_path, "temperature": self.temperature}
@property
def _llm_type(self):
return "custom_huggingface"
main_directory = os.path.dirname(os.path.abspath(__file__))
st.title("Web Page search Bot: Research Tool 📈")
st.sidebar.title("Article URLs")
urls = []
for i in range(3):
url = st.sidebar.text_input(f"URL {i+1}")
urls.append(url)
process_url_clicked = st.sidebar.button("Process URLs")
file_path_faiss = "faiss_store.pkl"
main_placeholder = st.empty()
# Load a pre-trained embedding model
embedding_model = HuggingFaceEmbeddings(model_name='sentence-transformers/all-MiniLM-L6-v2')
llm = CustomHuggingFaceLLM(model_name="meta-llama/Meta-Llama-3.1-8B", temperature=0.6)
if process_url_clicked:
# load data
loader = UnstructuredURLLoader(urls=urls)
main_placeholder.text("Data Loading...Started...✅✅✅")
data = loader.load()
# split data
# Do not include unnecessary separators like , and . It will reduce chunks too small.
text_splitter = RecursiveCharacterTextSplitter(
separators=['\n\n'],
chunk_size=1000,
chunk_overlap=100
)
main_placeholder.text("Text Splitter...Started...✅✅✅")
docs = text_splitter.split_documents(data)
# create embeddings and save it to FAISS index
vectorstore_faiss = FAISS.from_documents(documents=docs,embedding=embedding_model)
main_placeholder.text("Embedding Vector Started Building...✅✅✅")
time.sleep(2)
# Save the FAISS index to a pickle file
with open(file_path_faiss, "wb") as f:
pickle.dump(vectorstore_faiss, f)
query = main_placeholder.text_input("Question: ")
if query:
if os.path.exists(file_path_faiss):
with open(file_path_faiss, "rb") as f:
vectorstore = pickle.load(f)
chain = RetrievalQAWithSourcesChain.from_llm(llm=llm, retriever=vectorstore.as_retriever(), verbose=True) # type: ignore
result = chain({"question": query}, return_only_outputs=True)
# result will be a dictionary of this format --> {"answer": "", "sources": [] }
st.header("Answer")
st.write(result["answer"])
# Display sources, if available
sources = result.get("sources", "")
if sources:
st.subheader("Sources:")
sources_list = sources.split("\n") # Split the sources by newline
for source in sources_list:
st.write(source) |