Carlos Salgado commited on
Commit
a565cf7
·
unverified ·
2 Parent(s): b5a892f 6a820a4

Merge branch 'main' into main

Browse files
Files changed (3) hide show
  1. Dockerfile +1 -1
  2. app.py +140 -32
  3. app_V2.py +247 -0
Dockerfile CHANGED
@@ -25,7 +25,7 @@ COPY backend .
25
 
26
  # Install backend dependencies
27
  COPY backend/requirements.txt .
28
- RUN pip install --no-cache-dir -r requirements.txt --vvv
29
 
30
  # Stage 3: Serve frontend and backend using nginx and gunicorn
31
  FROM nginx:latest AS production
 
25
 
26
  # Install backend dependencies
27
  COPY backend/requirements.txt .
28
+ RUN pip install --no-cache-dir -r requirements.txt
29
 
30
  # Stage 3: Serve frontend and backend using nginx and gunicorn
31
  FROM nginx:latest AS production
app.py CHANGED
@@ -1,43 +1,151 @@
1
- import io
2
  import os
3
- import streamlit as st
4
- import tempfile
 
 
 
 
 
 
 
 
 
 
 
 
 
5
 
6
- from scripts import analyze_metadata, generate_metadata, ingest, MODEL_NAME
7
 
8
- st.title('# DocVerifyRAG')
9
- st.write('## Anomaly detection for BIM document metadata')
10
 
11
- with st.form('analyze_form'):
12
- st.write('Enter your file metadata in the following schema:')
13
- text = st.text_input(label='Filename, Description, Discipline',
14
- value="", placeholder=str)
15
- submitted = st.form_submit_button('Submit')
16
 
17
- if submitted:
18
- filename, description, discipline = text.split(',')
 
19
 
20
- st.write('## Analyzing with Vectara + together.ai')
21
- analysis = analyze_metadata(filename, description, discipline)
22
 
23
- st.write(analysis)
 
 
24
 
25
- st.write('## Generate metadata?')
26
- uploaded_file = st.file_uploader("Choose a PDF file", type=["pdf","txt"])
27
 
28
- if uploaded_file is not None:
29
- with tempfile.NamedTemporaryFile(delete=False, suffix=os.path.splitext(uploaded_file.name)[1]) as tmp:
30
- tmp.write(uploaded_file.read())
31
- file_path = tmp.name
32
- st.write(f'Created temporary file {file_path}')
33
 
34
- docs = ingest(file_path)
35
- st.write('## Querying Together.ai API')
36
- metadata = generate_metadata(docs)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
37
 
38
- form = st.form(key='generate_form')
39
- st.write(f'## Suggested Metadata Generated by {MODEL_NAME}')
40
- st.write(f'### {metadata}')
41
- delete_file_button = form.form_submit_button(label='Delete file')
42
- if delete_file_button:
43
- os.remove(file_path)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  import os
2
+ import io
3
+ import argparse
4
+ import json
5
+ import openai
6
+ import sys
7
+ from dotenv import load_dotenv
8
+ from langchain_community.document_loaders import TextLoader
9
+ from langchain_community.document_loaders import UnstructuredPDFLoader
10
+ from langchain_community.embeddings import HuggingFaceEmbeddings
11
+ from langchain_community.vectorstores import Vectara
12
+ from langchain_core.output_parsers import StrOutputParser
13
+ from langchain_core.prompts import ChatPromptTemplate
14
+ from langchain_core.runnables import RunnablePassthrough
15
+ from langchain.prompts import PromptTemplate
16
+ from langchain_text_splitters import RecursiveCharacterTextSplitter
17
 
 
18
 
19
+ load_dotenv()
 
20
 
21
+ MODEL_NAME = "mistralai/Mixtral-8x7B-Instruct-v0.1"
 
 
 
 
22
 
23
+ vectara_customer_id = os.environ['VECTARA_CUSTOMER_ID']
24
+ vectara_corpus_id = os.environ['VECTARA_CORPUS_ID']
25
+ vectara_api_key = os.environ['VECTARA_API_KEY']
26
 
27
+ embeddings = HuggingFaceEmbeddings(model_name="intfloat/multilingual-e5-large")
 
28
 
29
+ vectara = Vectara(vectara_customer_id=vectara_customer_id,
30
+ vectara_corpus_id=vectara_corpus_id,
31
+ vectara_api_key=vectara_api_key)
32
 
 
 
33
 
34
+ summary_config = {"is_enabled": True, "max_results": 3, "response_lang": "eng"}
35
+ retriever = vectara.as_retriever(
36
+ search_kwargs={"k": 3, "summary_config": summary_config}
37
+ )
 
38
 
39
+ template = """
40
+ passage: You are a helpful assistant that understands BIM building documents.
41
+ passage: You will analyze BIM document metadata composed of filename, description, and engineering discipline.
42
+ passage: The metadata is written in German.
43
+ passage: Filename: {filename}, Description: {description}, Engineering discipline: {discipline}.
44
+ query: Does the filename match other filenames within the same discipline?
45
+ query: Does the description match the engineering discipline?
46
+ query: How different is the metadata to your curated information?
47
+ query: Highligh any discrepancies and comment on wether or not the metadata is anomalous.
48
+ """
49
+
50
+ prompt = PromptTemplate(template=template, input_variables=['filename', 'description', 'discipline'])
51
+
52
+
53
+ def get_sources(documents):
54
+ return documents[:-1]
55
+
56
+ def get_summary(documents):
57
+ return documents[-1].page_content
58
+
59
+ def ingest(file_path):
60
+ extension = os.path.splitext(file_path)[1].lower()
61
+
62
+ if extension == '.pdf':
63
+ loader = UnstructuredPDFLoader(file_path)
64
+ elif extension == '.txt':
65
+ loader = TextLoader(file_path)
66
+ else:
67
+ raise NotImplementedError('Only .txt or .pdf files are supported')
68
+
69
+ # transform locally
70
+ documents = loader.load()
71
+ text_splitter = RecursiveCharacterTextSplitter(chunk_size=1000, chunk_overlap=0,
72
+ separators=[
73
+ "\n\n",
74
+ "\n",
75
+ " ",
76
+ ",",
77
+ "\uff0c", # Fullwidth comma
78
+ "\u3001", # Ideographic comma
79
+ "\uff0e", # Fullwidth full stop
80
+ # "\u200B", # Zero-width space (Asian languages)
81
+ # "\u3002", # Ideographic full stop (Asian languages)
82
+ "",
83
+ ])
84
+ docs = text_splitter.split_documents(documents)
85
+
86
+ return docs
87
+
88
+
89
+
90
+ def generate_metadata(docs):
91
+ prompt_template = """
92
+ BimDiscipline = ['plumbing', 'network', 'heating', 'electrical', 'ventilation', 'architecture']
93
 
94
+ You are a helpful assistant that understands BIM documents and engineering disciplines. Your answer should be in JSON format and only include the filename, a short description, and the engineering discipline the document belongs to, distinguishing between {[d.value for d in BimDiscipline]} based on the given document."
95
+
96
+ Analyze the provided document, which could be in either German or English. Extract the filename, its description, and infer the engineering discipline it belongs to. Document:
97
+ context="
98
+ """
99
+ # plain text
100
+ filepath = [doc.metadata for doc in docs][0]['source']
101
+ context = "".join(
102
+ [doc.page_content.replace('\n\n','').replace('..','') for doc in docs])
103
+
104
+ prompt = f'{prompt_template}{context}"\nFilepath:{filepath}'
105
+
106
+ #print(prompt)
107
+
108
+ # Create client
109
+ client = openai.OpenAI(
110
+ base_url="https://api.together.xyz/v1",
111
+ api_key=os.environ["TOGETHER_API_KEY"],
112
+ #api_key=userdata.get('TOGETHER_API_KEY'),
113
+ )
114
+
115
+ # Call the LLM with the JSON schema
116
+ chat_completion = client.chat.completions.create(
117
+ model=MODEL_NAME,
118
+ messages=[
119
+ {
120
+ "role": "system",
121
+ "content": f"You are a helpful assistant that responsds in JSON format"
122
+ },
123
+ {
124
+ "role": "user",
125
+ "content": prompt
126
+ }
127
+ ]
128
+ )
129
+
130
+ return json.loads(chat_completion.choices[0].message.content)
131
+
132
+
133
+ def analyze_metadata(filename, description, discipline):
134
+ formatted_prompt = prompt.format(filename=filename, description=description, discipline=discipline)
135
+ return (retriever | get_summary).invoke(formatted_prompt)
136
+
137
+
138
+ if __name__ == "__main__":
139
+ parser = argparse.ArgumentParser(description="Generate metadata for a BIM document")
140
+ parser.add_argument("document", metavar="FILEPATH", type=str,
141
+ help="Path to the BIM document")
142
+
143
+ args = parser.parse_args()
144
+
145
+ if not os.path.exists(args.document) or not os.path.isfile(args.document):
146
+ print("File '{}' not found or not accessible.".format(args.document))
147
+ sys.exit(-1)
148
+
149
+ docs = ingest(args.document)
150
+ metadata = generate_metadata(docs)
151
+ print(metadata)
app_V2.py ADDED
@@ -0,0 +1,247 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import tempfile
2
+ import streamlit as st
3
+ from PyPDF2 import PdfReader
4
+ from langchain.text_splitter import CharacterTextSplitter
5
+ from langchain.embeddings import OpenAIEmbeddings
6
+ from langchain.vectorstores import FAISS
7
+ from langchain.chat_models import ChatOpenAI
8
+ from langchain.memory import ConversationBufferMemory
9
+ from langchain.chains import ConversationalRetrievalChain
10
+ import os
11
+ import pickle
12
+ from datetime import datetime
13
+ from backend.generate_metadata import generate_metadata, ingest
14
+
15
+ MODEL_NAME = "mixtral"
16
+ css = '''
17
+ <style>
18
+ .chat-message {
19
+ padding: 1.5rem; border-radius: 0.5rem; margin-bottom: 1rem; display: flex
20
+ }
21
+ .chat-message.user {
22
+ background-color: #2b313e
23
+ }
24
+ .chat-message.bot {
25
+ background-color: #475063
26
+ }
27
+ .chat-message .avatar {
28
+ width: 20%;
29
+ }
30
+ .chat-message .avatar img {
31
+ max-width: 78px;
32
+ max-height: 78px;
33
+ border-radius: 50%;
34
+ object-fit: cover;
35
+ }
36
+ .chat-message .message {
37
+ width: 80%;
38
+ padding: 0 1.5rem;
39
+ color: #fff;
40
+ }
41
+ '''
42
+ bot_template = '''
43
+ <div class="chat-message bot">
44
+ <div class="avatar">
45
+ <img src="https://i.ibb.co/cN0nmSj/Screenshot-2023-05-28-at-02-37-21.png"
46
+ style="max-height: 78px; max-width: 78px; border-radius: 50%; object-fit: cover;">
47
+ </div>
48
+ <div class="message">{{MSG}}</div>
49
+ </div>
50
+ '''
51
+ user_template = '''
52
+ <div class="chat-message user">
53
+ <div class="avatar">
54
+ <img src="https://i.ibb.co/rdZC7LZ/Photo-logo-1.png">
55
+ </div>
56
+ <div class="message">{{MSG}}</div>
57
+ </div>
58
+ '''
59
+
60
+
61
+ def get_pdf_text(pdf_docs):
62
+ text = ""
63
+ for pdf in pdf_docs:
64
+ pdf_reader = PdfReader(pdf)
65
+ for page in pdf_reader.pages:
66
+ text += page.extract_text()
67
+ return text
68
+
69
+
70
+ def get_text_chunks(text):
71
+ text_splitter = CharacterTextSplitter(
72
+ separator="\n",
73
+ chunk_size=1000,
74
+ chunk_overlap=200,
75
+ length_function=len
76
+ )
77
+ chunks = text_splitter.split_text(text)
78
+ return chunks
79
+
80
+
81
+ def get_vectorstore(text_chunks):
82
+ embeddings = OpenAIEmbeddings()
83
+ # embeddings = HuggingFaceInstructEmbeddings(model_name="hkunlp/instructor-xl")
84
+ vectorstore = FAISS.from_texts(texts=text_chunks, embedding=embeddings)
85
+ return vectorstore
86
+
87
+
88
+ def get_conversation_chain(vectorstore):
89
+ llm = ChatOpenAI()
90
+ # llm = HuggingFaceHub(repo_id="google/flan-t5-xxl", model_kwargs={"temperature":0.5, "max_length":512})
91
+
92
+ memory = ConversationBufferMemory(
93
+ memory_key='chat_history', return_messages=True)
94
+ conversation_chain = ConversationalRetrievalChain.from_llm(
95
+ llm=llm,
96
+ retriever=vectorstore.as_retriever(),
97
+ memory=memory
98
+ )
99
+ return conversation_chain
100
+
101
+
102
+ def handle_userinput(user_question):
103
+ response = st.session_state.conversation({'question': user_question})
104
+ st.session_state.chat_history = response['chat_history']
105
+
106
+ for i, message in enumerate(st.session_state.chat_history):
107
+ # Display user message
108
+ if i % 2 == 0:
109
+ st.write(user_template.replace("{{MSG}}", message.content), unsafe_allow_html=True)
110
+ else:
111
+ print(message)
112
+ # Display AI response
113
+ st.write(bot_template.replace("{{MSG}}", message.content), unsafe_allow_html=True)
114
+
115
+
116
+ def safe_vec_store():
117
+ # USE VECTARA INSTEAD
118
+ os.makedirs('vectorstore', exist_ok=True)
119
+ filename = 'vectors' + datetime.now().strftime('%Y%m%d%H%M') + '.pkl'
120
+ file_path = os.path.join('vectorstore', filename)
121
+ vector_store = st.session_state.vectorstore
122
+
123
+ # Serialize and save the entire FAISS object using pickle
124
+ with open(file_path, 'wb') as f:
125
+ pickle.dump(vector_store, f)
126
+
127
+
128
+ """
129
+ def main():
130
+
131
+
132
+
133
+ st.subheader("Your documents")
134
+
135
+ if st.session_state.classify:
136
+ pdf_doc = st.file_uploader("Upload your PDFs here and click on 'Process'", accept_multiple_files=False)
137
+ else:
138
+ pdf_docs = st.file_uploader("Upload your PDFs here and click on 'Process'", accept_multiple_files=True)
139
+ filenames = [file.name for file in pdf_docs if file is not None]
140
+ if st.button("Process"):
141
+ with st.spinner("Processing"):
142
+ if st.session_state.classify:
143
+ # THE CLASSIFICATION APP
144
+ st.write("Classifying")
145
+ plain_text_doc = ingest(pdf_doc.name)
146
+ classification_result = generate_metadata(plain_text_doc)
147
+ st.write(classification_result)
148
+ else:
149
+ # NORMAL RAG
150
+ loaded_vec_store = None
151
+ for filename in filenames:
152
+ if ".pkl" in filename:
153
+ file_path = os.path.join('vectorstore', filename)
154
+ with open(file_path, 'rb') as f:
155
+ loaded_vec_store = pickle.load(f)
156
+ raw_text = get_pdf_text(pdf_docs)
157
+ text_chunks = get_text_chunks(raw_text)
158
+ vec = get_vectorstore(text_chunks)
159
+ if loaded_vec_store:
160
+ vec.merge_from(loaded_vec_store)
161
+ st.warning("loaded vectorstore")
162
+ if "vectorstore" in st.session_state:
163
+ vec.merge_from(st.session_state.vectorstore)
164
+ st.warning("merged to existing")
165
+ st.session_state.vectorstore = vec
166
+ st.session_state.conversation = get_conversation_chain(vec)
167
+ st.success("data loaded")
168
+
169
+ if "conversation" not in st.session_state:
170
+ st.session_state.conversation = None
171
+ if "chat_history" not in st.session_state:
172
+ st.session_state.chat_history = None
173
+
174
+ user_question = st.text_input("Ask a question about your documents:")
175
+ if user_question:
176
+ handle_userinput(user_question)
177
+ with st.sidebar:
178
+ st.subheader("Classification instructions")
179
+ classifier_docs = st.file_uploader("Upload your instructions here and click on 'Process'",
180
+ accept_multiple_files=True)
181
+ filenames = [file.name for file in classifier_docs if file is not None]
182
+
183
+ if st.button("Process Classification"):
184
+ st.session_state.classify = True
185
+ with st.spinner("Processing"):
186
+ st.warning("set classify")
187
+ time.sleep(3)
188
+
189
+ if st.button("Save Embeddings"):
190
+ if "vectorstore" in st.session_state:
191
+ safe_vec_store()
192
+ # st.session_state.vectorstore.save_local("faiss_index")
193
+ st.sidebar.success("saved")
194
+ else:
195
+ st.sidebar.warning("No embeddings to save. Please process documents first.")
196
+
197
+ if st.button("Load Embeddings"):
198
+ st.warning("this function is not in use, just upload the vectorstore")
199
+ """
200
+
201
+
202
+ def main():
203
+
204
+ st.set_page_config(page_title="Doc Verify RAG", page_icon=":mag:")
205
+ st.write('Anomaly detection for document metadata', unsafe_allow_html=True)
206
+ st.header("Doc Verify RAG :mag:")
207
+
208
+ def set_pw():
209
+ st.session_state.openai_api_key = True
210
+
211
+ if "openai_api_key" not in st.session_state:
212
+ st.session_state.openai_api_key = False
213
+ if "openai_org" not in st.session_state:
214
+ st.session_state.openai_org = False
215
+ if "classify" not in st.session_state:
216
+ st.session_state.classify = False
217
+
218
+ col1, col2 = st.columns(2)
219
+ with col1:
220
+ uploaded_file = st.file_uploader("Choose a PDF file", type=["pdf", "txt"])
221
+
222
+ if uploaded_file is not None:
223
+ try:
224
+ with tempfile.NamedTemporaryFile(delete=False, suffix=os.path.splitext(uploaded_file.name)[1]) as tmp:
225
+ tmp.write(uploaded_file.read())
226
+ file_path = tmp.name
227
+ st.write(f'Created temporary file {file_path}')
228
+
229
+ docs = ingest(file_path)
230
+ st.write('## Querying Together.ai API')
231
+ metadata = generate_metadata(docs)
232
+ st.write(f'## Metadata Generated by {MODEL_NAME}')
233
+ st.write(metadata)
234
+
235
+ # Clean up the temporary file
236
+ os.remove(file_path)
237
+
238
+ except Exception as e:
239
+ st.error(f'Error: {e}')
240
+ with col2:
241
+ OPENAI_API_KEY = st.text_input("OPENAI API KEY:", type="password",
242
+ disabled=st.session_state.openai_api_key, on_change=set_pw)
243
+ classification = st.file_uploader("upload the metadata", type=["csv", "txt"])
244
+
245
+
246
+ if __name__ == '__main__':
247
+ main()