Create app.py
Browse files
app.py
ADDED
@@ -0,0 +1,119 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from flask import Flask, request, jsonify
|
2 |
+
import os
|
3 |
+
from langchain_community.document_loaders import PyPDFLoader
|
4 |
+
from langchain.text_splitter import RecursiveCharacterTextSplitter
|
5 |
+
from langchain_community.vectorstores import Chroma
|
6 |
+
from langchain.chains import ConversationalRetrievalChain
|
7 |
+
from langchain_community.embeddings import HuggingFaceEmbeddings
|
8 |
+
from langchain_community.llms import HuggingFaceEndpoint
|
9 |
+
from langchain.memory import ConversationBufferMemory
|
10 |
+
from pathlib import Path
|
11 |
+
import chromadb
|
12 |
+
from unidecode import unidecode
|
13 |
+
import re
|
14 |
+
|
15 |
+
app = Flask(__name__)
|
16 |
+
|
17 |
+
# Configuration variables
|
18 |
+
PDF_PATH = "path/to/your/static.pdf" # Replace with your static PDF path
|
19 |
+
CHUNK_SIZE = 512
|
20 |
+
CHUNK_OVERLAP = 24
|
21 |
+
LLM_MODEL = "mistralai/Mistral-7B-Instruct-v0.2"
|
22 |
+
TEMPERATURE = 0.1
|
23 |
+
MAX_TOKENS = 512
|
24 |
+
TOP_K = 20
|
25 |
+
|
26 |
+
# Load PDF document and create doc splits
|
27 |
+
def load_doc(pdf_path, chunk_size, chunk_overlap):
|
28 |
+
loader = PyPDFLoader(pdf_path)
|
29 |
+
pages = loader.load()
|
30 |
+
text_splitter = RecursiveCharacterTextSplitter(chunk_size=chunk_size, chunk_overlap=chunk_overlap)
|
31 |
+
doc_splits = text_splitter.split_documents(pages)
|
32 |
+
return doc_splits
|
33 |
+
|
34 |
+
# Create vector database
|
35 |
+
def create_db(splits, collection_name):
|
36 |
+
embedding = HuggingFaceEmbeddings()
|
37 |
+
new_client = chromadb.EphemeralClient()
|
38 |
+
vectordb = Chroma.from_documents(
|
39 |
+
documents=splits,
|
40 |
+
embedding=embedding,
|
41 |
+
client=new_client,
|
42 |
+
collection_name=collection_name,
|
43 |
+
)
|
44 |
+
return vectordb
|
45 |
+
|
46 |
+
# Initialize langchain LLM chain
|
47 |
+
def initialize_llmchain(llm_model, temperature, max_tokens, top_k, vector_db):
|
48 |
+
llm = HuggingFaceEndpoint(
|
49 |
+
repo_id=llm_model,
|
50 |
+
temperature=temperature,
|
51 |
+
max_new_tokens=max_tokens,
|
52 |
+
top_k=top_k,
|
53 |
+
)
|
54 |
+
|
55 |
+
memory = ConversationBufferMemory(
|
56 |
+
memory_key="chat_history",
|
57 |
+
output_key='answer',
|
58 |
+
return_messages=True
|
59 |
+
)
|
60 |
+
retriever = vector_db.as_retriever()
|
61 |
+
qa_chain = ConversationalRetrievalChain.from_llm(
|
62 |
+
llm,
|
63 |
+
retriever=retriever,
|
64 |
+
chain_type="stuff",
|
65 |
+
memory=memory,
|
66 |
+
return_source_documents=True,
|
67 |
+
verbose=False,
|
68 |
+
)
|
69 |
+
return qa_chain
|
70 |
+
|
71 |
+
# Generate collection name for vector database
|
72 |
+
def create_collection_name(filepath):
|
73 |
+
collection_name = Path(filepath).stem
|
74 |
+
collection_name = collection_name.replace(" ", "-")
|
75 |
+
collection_name = unidecode(collection_name)
|
76 |
+
collection_name = re.sub('[^A-Za-z0-9]+', '-', collection_name)
|
77 |
+
collection_name = collection_name[:50]
|
78 |
+
if len(collection_name) < 3:
|
79 |
+
collection_name = collection_name + 'xyz'
|
80 |
+
if not collection_name[0].isalnum():
|
81 |
+
collection_name = 'A' + collection_name[1:]
|
82 |
+
if not collection_name[-1].isalnum():
|
83 |
+
collection_name = collection_name[:-1] + 'Z'
|
84 |
+
return collection_name
|
85 |
+
|
86 |
+
# Initialize database and QA chain
|
87 |
+
doc_splits = load_doc(PDF_PATH, CHUNK_SIZE, CHUNK_OVERLAP)
|
88 |
+
collection_name = create_collection_name(PDF_PATH)
|
89 |
+
vector_db = create_db(doc_splits, collection_name)
|
90 |
+
qa_chain = initialize_llmchain(LLM_MODEL, TEMPERATURE, MAX_TOKENS, TOP_K, vector_db)
|
91 |
+
|
92 |
+
@app.route('/chat', methods=['POST'])
|
93 |
+
def chat():
|
94 |
+
data = request.json
|
95 |
+
message = data.get('message', '')
|
96 |
+
history = data.get('history', [])
|
97 |
+
|
98 |
+
formatted_chat_history = []
|
99 |
+
for user_message, bot_message in history:
|
100 |
+
formatted_chat_history.append(f"User: {user_message}")
|
101 |
+
formatted_chat_history.append(f"Assistant: {bot_message}")
|
102 |
+
|
103 |
+
response = qa_chain({"question": message, "chat_history": formatted_chat_history})
|
104 |
+
response_answer = response["answer"]
|
105 |
+
if response_answer.find("Helpful Answer:") != -1:
|
106 |
+
response_answer = response_answer.split("Helpful Answer:")[-1]
|
107 |
+
response_sources = response["source_documents"]
|
108 |
+
|
109 |
+
result = {
|
110 |
+
"answer": response_answer,
|
111 |
+
"sources": [
|
112 |
+
{"content": doc.page_content.strip(), "page": doc.metadata["page"] + 1}
|
113 |
+
for doc in response_sources
|
114 |
+
]
|
115 |
+
}
|
116 |
+
return jsonify(result)
|
117 |
+
|
118 |
+
if __name__ == '__main__':
|
119 |
+
app.run(debug=True, host='0.0.0.0', port=5000)
|