Spaces:
Sleeping
Sleeping
Create app.py
Browse files
app.py
ADDED
@@ -0,0 +1,171 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import streamlit as st
|
2 |
+
from dotenv import load_dotenv
|
3 |
+
import os
|
4 |
+
from htmlTemplate import css, bot_template, user_template
|
5 |
+
import PyPDF2
|
6 |
+
from langchain.text_splitter import RecursiveCharacterTextSplitter
|
7 |
+
from langchain_community.embeddings.spacy_embeddings import SpacyEmbeddings
|
8 |
+
from langchain_community.llms import LlamaCpp
|
9 |
+
from langchain.embeddings import HuggingFaceEmbeddings
|
10 |
+
from langchain.vectorstores import FAISS
|
11 |
+
from langchain.memory import ConversationBufferMemory
|
12 |
+
from langchain.chains import ConversationalRetrievalChain
|
13 |
+
from langchain.prompts import PromptTemplate
|
14 |
+
from sentence_transformers import SentenceTransformer, util
|
15 |
+
from langchain_openai import AzureOpenAIEmbeddings
|
16 |
+
from langchain_openai import OpenAIEmbeddings
|
17 |
+
from langchain_community.embeddings.fastembed import FastEmbedEmbeddings
|
18 |
+
from langchain_openai import ChatOpenAI
|
19 |
+
os.environ["OPENAI_API_KEY"] = "sk-.............."
|
20 |
+
os.environ["GROQ_API_KEY"]="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"
|
21 |
+
from langchain_groq import ChatGroq
|
22 |
+
|
23 |
+
llmtemplate = """You’re an AI information specialist with a strong emphasis on extracting accurate information from markdown documents. Your expertise involves summarizing data succinctly while adhering to strict guidelines about neutrality and clarity.
|
24 |
+
|
25 |
+
Your task is to answer a specific question based on a provided markdown document. Here is the question you need to address:
|
26 |
+
{question}
|
27 |
+
|
28 |
+
Keep in mind the following instructions:
|
29 |
+
- Your response should be direct and factual, limited to 50 words and 2-3 sentences.
|
30 |
+
- Avoid using introductory phrases like "yes" or "no."
|
31 |
+
- Maintain an ethical and unbiased tone, steering clear of harmful or offensive content.
|
32 |
+
- If the document lacks relevant information, respond with "I cannot provide an answer based on the provided document."
|
33 |
+
- Do not fabricate information, include questions, or use confirmatory phrases.
|
34 |
+
- Remember not to prompt for additional information or ask any questions.
|
35 |
+
|
36 |
+
Ensure your response is strictly based on the content of the markdown document.
|
37 |
+
"""
|
38 |
+
|
39 |
+
|
40 |
+
|
41 |
+
def prepare_docs(pdf_docs):
|
42 |
+
docs = []
|
43 |
+
metadata = []
|
44 |
+
content = []
|
45 |
+
|
46 |
+
for pdf in pdf_docs:
|
47 |
+
print(pdf.name)
|
48 |
+
pdf_reader = PyPDF2.PdfReader(pdf)
|
49 |
+
for index, text in enumerate(pdf_reader.pages):
|
50 |
+
doc_page = {'title': pdf.name + " page " + str(index + 1),
|
51 |
+
'content': pdf_reader.pages[index].extract_text()}
|
52 |
+
docs.append(doc_page)
|
53 |
+
for doc in docs:
|
54 |
+
content.append(doc["content"])
|
55 |
+
metadata.append({
|
56 |
+
"title": doc["title"]
|
57 |
+
})
|
58 |
+
return content, metadata
|
59 |
+
|
60 |
+
|
61 |
+
def get_text_chunks(content, metadata):
|
62 |
+
text_splitter = RecursiveCharacterTextSplitter.from_tiktoken_encoder(
|
63 |
+
chunk_size=1024,
|
64 |
+
chunk_overlap=256,
|
65 |
+
)
|
66 |
+
split_docs = text_splitter.create_documents(content, metadatas=metadata)
|
67 |
+
print(f"Split documents into {len(split_docs)} passages")
|
68 |
+
return split_docs
|
69 |
+
|
70 |
+
|
71 |
+
def ingest_into_vectordb(split_docs):
|
72 |
+
# embeddings = OpenAIEmbeddings()
|
73 |
+
# embeddings = FastEmbedEmbeddings()
|
74 |
+
embeddings = SpacyEmbeddings(model_name="en_core_web_sm")
|
75 |
+
db = FAISS.from_documents(split_docs, embeddings)
|
76 |
+
DB_FAISS_PATH = 'vectorstore/db_faiss'
|
77 |
+
db.save_local(DB_FAISS_PATH)
|
78 |
+
return db
|
79 |
+
|
80 |
+
|
81 |
+
def get_conversation_chain(vectordb):
|
82 |
+
# llama_llm = ChatOpenAI(temperature=0.7, model="gpt-3.5-turbo")
|
83 |
+
llm = ChatGroq(model="llama3-70b-8192", temperature=0.25)
|
84 |
+
retriever = vectordb.as_retriever()
|
85 |
+
CONDENSE_QUESTION_PROMPT = PromptTemplate.from_template(llmtemplate)
|
86 |
+
|
87 |
+
memory = ConversationBufferMemory(
|
88 |
+
memory_key='chat_history', return_messages=True, output_key='answer')
|
89 |
+
|
90 |
+
conversation_chain = (ConversationalRetrievalChain.from_llm
|
91 |
+
(llm=llm,
|
92 |
+
retriever=retriever,
|
93 |
+
#condense_question_prompt=CONDENSE_QUESTION_PROMPT,
|
94 |
+
memory=memory,
|
95 |
+
return_source_documents=True))
|
96 |
+
print("Conversational Chain created for the LLM using the vector store")
|
97 |
+
return conversation_chain
|
98 |
+
|
99 |
+
def validate_answer_against_sources(response_answer, source_documents):
|
100 |
+
model = SentenceTransformer('all-MiniLM-L6-v2')
|
101 |
+
similarity_threshold = 0.5
|
102 |
+
source_texts = [doc.page_content for doc in source_documents]
|
103 |
+
|
104 |
+
answer_embedding = model.encode(response_answer, convert_to_tensor=True)
|
105 |
+
source_embeddings = model.encode(source_texts, convert_to_tensor=True)
|
106 |
+
|
107 |
+
cosine_scores = util.pytorch_cos_sim(answer_embedding, source_embeddings)
|
108 |
+
|
109 |
+
|
110 |
+
if any(score.item() > similarity_threshold for score in cosine_scores[0]):
|
111 |
+
return True
|
112 |
+
|
113 |
+
return False
|
114 |
+
|
115 |
+
def handle_userinput(user_question):
|
116 |
+
response = st.session_state.conversation({'question': user_question})
|
117 |
+
st.session_state.chat_history = response['chat_history']
|
118 |
+
|
119 |
+
for i, message in enumerate(st.session_state.chat_history):
|
120 |
+
print(i)
|
121 |
+
if i % 2 == 0:
|
122 |
+
st.write(user_template.replace(
|
123 |
+
"{{MSG}}", message.content), unsafe_allow_html=True)
|
124 |
+
else:
|
125 |
+
print(message.content)
|
126 |
+
st.write(bot_template.replace(
|
127 |
+
"{{MSG}}", message.content), unsafe_allow_html=True)
|
128 |
+
|
129 |
+
|
130 |
+
|
131 |
+
def main():
|
132 |
+
load_dotenv()
|
133 |
+
|
134 |
+
st.set_page_config(page_title="Chat with your PDFs",
|
135 |
+
page_icon=":books:")
|
136 |
+
st.write(css, unsafe_allow_html=True)
|
137 |
+
|
138 |
+
if "conversation" not in st.session_state:
|
139 |
+
st.session_state.conversation = None
|
140 |
+
if "chat_history" not in st.session_state:
|
141 |
+
st.session_state.chat_history = []
|
142 |
+
|
143 |
+
st.header("Chat with multiple PDFs :books:")
|
144 |
+
user_question = st.text_input("Ask a question about your documents:")
|
145 |
+
|
146 |
+
if user_question:
|
147 |
+
handle_userinput(user_question)
|
148 |
+
|
149 |
+
with st.sidebar:
|
150 |
+
st.subheader("Your documents")
|
151 |
+
pdf_docs = st.file_uploader(
|
152 |
+
"Upload your PDFs here and click on 'Process'", accept_multiple_files=True)
|
153 |
+
|
154 |
+
if st.button("Process"):
|
155 |
+
with st.spinner("Processing"):
|
156 |
+
# get pdf text
|
157 |
+
content, metadata = prepare_docs(pdf_docs)
|
158 |
+
|
159 |
+
# get the text chunks
|
160 |
+
split_docs = get_text_chunks(content, metadata)
|
161 |
+
|
162 |
+
# create vector store
|
163 |
+
vectorstore = ingest_into_vectordb(split_docs)
|
164 |
+
|
165 |
+
# create conversation chain
|
166 |
+
st.session_state.conversation = get_conversation_chain(
|
167 |
+
vectorstore)
|
168 |
+
|
169 |
+
|
170 |
+
if __name__ == '__main__':
|
171 |
+
main()
|