Spaces:
Sleeping
Sleeping
Create app.py
Browse files
app.py
ADDED
@@ -0,0 +1,188 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Install necessary packages
|
2 |
+
#!pip install streamlit
|
3 |
+
#!pip install wikipedia
|
4 |
+
#!pip install langchain_community
|
5 |
+
#!pip install sentence-transformers
|
6 |
+
#!pip install chromadb
|
7 |
+
#!pip install huggingface_hub
|
8 |
+
#!pip install transformers
|
9 |
+
|
10 |
+
import streamlit as st
|
11 |
+
from langchain_community.document_loaders import PyPDFLoader
|
12 |
+
from langchain.text_splitter import RecursiveCharacterTextSplitter, SentenceTransformersTokenTextSplitter
|
13 |
+
import chromadb
|
14 |
+
from chromadb.utils.embedding_functions import SentenceTransformerEmbeddingFunction
|
15 |
+
from huggingface_hub import login, InferenceClient
|
16 |
+
from sentence_transformers import CrossEncoder
|
17 |
+
import numpy as np
|
18 |
+
import random
|
19 |
+
import string
|
20 |
+
import os
|
21 |
+
|
22 |
+
# User variables
|
23 |
+
uploaded_file = st.sidebar.file_uploader("Upload your PDF", type="pdf")
|
24 |
+
model_name = 'mistralai/Mistral-7B-Instruct-v0.3'
|
25 |
+
HF_TOKEN = st.sidebar.text_input("Enter your Hugging Face token:", "", type="password")
|
26 |
+
|
27 |
+
|
28 |
+
# Initialize session state for error message
|
29 |
+
if 'error_message' not in st.session_state:
|
30 |
+
st.session_state.error_message = ""
|
31 |
+
|
32 |
+
# Function to validate token
|
33 |
+
def validate_token(token):
|
34 |
+
try:
|
35 |
+
# Attempt to log in with the provided token
|
36 |
+
login(token=token)
|
37 |
+
# Check if the token is valid by trying to access some data
|
38 |
+
HfApi().whoami()
|
39 |
+
return True
|
40 |
+
except Exception as e:
|
41 |
+
return False
|
42 |
+
|
43 |
+
# Validate the token and display appropriate message
|
44 |
+
if HF_TOKEN:
|
45 |
+
if validate_token(HF_TOKEN):
|
46 |
+
st.session_state.error_message = "" # Clear error message if the token is valid
|
47 |
+
st.sidebar.success("Token is valid!")
|
48 |
+
else:
|
49 |
+
st.session_state.error_message = "Invalid token. Please try again."
|
50 |
+
st.sidebar.error(st.session_state.error_message)
|
51 |
+
elif st.session_state.error_message:
|
52 |
+
st.sidebar.error(st.session_state.error_message)
|
53 |
+
|
54 |
+
if uploaded_file:
|
55 |
+
# Save the uploaded file temporarily
|
56 |
+
temp_file_path = os.path.join("/tmp", uploaded_file.name)
|
57 |
+
with open(temp_file_path, "wb") as f:
|
58 |
+
f.write(uploaded_file.getbuffer())
|
59 |
+
|
60 |
+
# Load the PDF using PyPDFLoader
|
61 |
+
docs = PyPDFLoader(temp_file_path).load()
|
62 |
+
|
63 |
+
# Clean up the temporary file after loading
|
64 |
+
os.remove(temp_file_path)
|
65 |
+
else:
|
66 |
+
st.warning("Please upload a PDF file.")
|
67 |
+
|
68 |
+
|
69 |
+
# Memory for chat history
|
70 |
+
if "history" not in st.session_state:
|
71 |
+
st.session_state.history = []
|
72 |
+
|
73 |
+
# Function to generate a random string for collection name
|
74 |
+
def generate_random_string(max_length=60):
|
75 |
+
if max_length > 60:
|
76 |
+
raise ValueError("The maximum length cannot exceed 60 characters.")
|
77 |
+
length = random.randint(1, max_length)
|
78 |
+
characters = string.ascii_letters + string.digits
|
79 |
+
return ''.join(random.choice(characters) for _ in range(length))
|
80 |
+
|
81 |
+
collection_name = generate_random_string()
|
82 |
+
|
83 |
+
# Function for query expansion
|
84 |
+
def augment_multiple_query(query):
|
85 |
+
client = InferenceClient(model_name, token=HF_TOKEN)
|
86 |
+
content = client.chat_completion(
|
87 |
+
messages=[
|
88 |
+
{
|
89 |
+
"role": "system",
|
90 |
+
"content": f"""You are a helpful assistant.
|
91 |
+
Suggest up to five additional related questions to help them find the information they need for the provided question.
|
92 |
+
Suggest only short questions without compound sentences. Suggest a variety of questions that cover different aspects of the topic.
|
93 |
+
Make sure they are complete questions, and that they are related to the original question."""
|
94 |
+
},
|
95 |
+
{
|
96 |
+
"role": "user",
|
97 |
+
"content": query
|
98 |
+
}
|
99 |
+
],
|
100 |
+
max_tokens=500,
|
101 |
+
)
|
102 |
+
return content.choices[0].message.content.split("\n")
|
103 |
+
|
104 |
+
# Function to handle RAG-based question answering
|
105 |
+
def rag_advanced(user_query):
|
106 |
+
|
107 |
+
|
108 |
+
# Text Splitting
|
109 |
+
character_splitter = RecursiveCharacterTextSplitter(separators=["\n\n", "\n", ". ", " ", ""], chunk_size=1000, chunk_overlap=0)
|
110 |
+
concat_texts = "".join([doc.page_content for doc in docs])
|
111 |
+
character_split_texts = character_splitter.split_text(concat_texts)
|
112 |
+
|
113 |
+
token_splitter = SentenceTransformersTokenTextSplitter(chunk_overlap=0, tokens_per_chunk=256)
|
114 |
+
token_split_texts = [text for text in character_split_texts for text in token_splitter.split_text(text)]
|
115 |
+
|
116 |
+
# Embedding and Document Storage
|
117 |
+
embedding_function = SentenceTransformerEmbeddingFunction()
|
118 |
+
chroma_client = chromadb.Client()
|
119 |
+
chroma_collection = chroma_client.create_collection(collection_name, embedding_function=embedding_function)
|
120 |
+
|
121 |
+
ids = [str(i) for i in range(len(token_split_texts))]
|
122 |
+
chroma_collection.add(ids=ids, documents=token_split_texts)
|
123 |
+
|
124 |
+
# Document Retrieval
|
125 |
+
augmented_queries = augment_multiple_query(user_query)
|
126 |
+
joint_query = [user_query] + augmented_queries
|
127 |
+
|
128 |
+
results = chroma_collection.query(query_texts=joint_query, n_results=5, include=['documents', 'embeddings'])
|
129 |
+
retrieved_documents = results['documents']
|
130 |
+
|
131 |
+
unique_documents = list(set(doc for docs in retrieved_documents for doc in docs))
|
132 |
+
|
133 |
+
# Re-Ranking
|
134 |
+
cross_encoder = CrossEncoder('cross-encoder/ms-marco-MiniLM-L-6-v2')
|
135 |
+
pairs = [[user_query, doc] for doc in unique_documents]
|
136 |
+
scores = cross_encoder.predict(pairs)
|
137 |
+
|
138 |
+
top_indices = np.argsort(scores)[::-1][:5]
|
139 |
+
top_documents = [unique_documents[idx] for idx in top_indices]
|
140 |
+
|
141 |
+
# LLM Reference
|
142 |
+
client = InferenceClient(model_name, token=HF_TOKEN)
|
143 |
+
response = ""
|
144 |
+
for message in client.chat_completion(
|
145 |
+
messages=[
|
146 |
+
{
|
147 |
+
"role": "system",
|
148 |
+
"content": f"""You are a helpful assitant.
|
149 |
+
You will be shown the user's questions, and the relevant information from the related documents.
|
150 |
+
Answer the user's question using only this information."""
|
151 |
+
},
|
152 |
+
{
|
153 |
+
"role": "user",
|
154 |
+
"content": f"Questions: {user_query}. \n Information: {top_documents}"
|
155 |
+
}
|
156 |
+
],
|
157 |
+
max_tokens=500,
|
158 |
+
stream=True,
|
159 |
+
):
|
160 |
+
response += message.choices[0].delta.content
|
161 |
+
|
162 |
+
return response
|
163 |
+
|
164 |
+
# Streamlit UI
|
165 |
+
st.title("PDF RAG Chatbot")
|
166 |
+
st.markdown("Upload your PDF and enter your 🤗 token!")
|
167 |
+
st.link_button("Get Token Here", "https://huggingface.co/settings/tokens")
|
168 |
+
|
169 |
+
# Input box for the user to type their message
|
170 |
+
if uploaded_file:
|
171 |
+
user_input = st.text_input("You: ", "", placeholder="Type your question here...")
|
172 |
+
if user_input:
|
173 |
+
response = rag_advanced(user_input)
|
174 |
+
st.session_state.history.append({"user": user_input, "bot": response})
|
175 |
+
|
176 |
+
# Display the conversation history
|
177 |
+
for chat in st.session_state.history:
|
178 |
+
st.write(f"You: {chat['user']}")
|
179 |
+
st.write(f"Bot: {chat['bot']}")
|
180 |
+
|
181 |
+
|
182 |
+
st.markdown("-----------------")
|
183 |
+
st.markdown("What is this app?")
|
184 |
+
st.markdown("""This is a simple RAG application using PDF import.
|
185 |
+
The model for chat is Mistral-7B-Instruct-v0.3.
|
186 |
+
Main libraries: Langchain (text splitting), Chromadb (vector store)
|
187 |
+
This RAG uses query expansion and re-ranking to improve the quality.
|
188 |
+
Feel free to check the files or DM me for any questions. Thank you.""")
|