afatuddin commited on
Commit
5342abb
·
verified ·
1 Parent(s): d1bedad

Upload app.py

Browse files
Files changed (1) hide show
  1. app.py +232 -0
app.py ADDED
@@ -0,0 +1,232 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import streamlit as st
2
+ from PyPDF2 import PdfReader
3
+ import os
4
+ from dotenv import load_dotenv
5
+
6
+ # Load environment variables from .env file
7
+ load_dotenv()
8
+
9
+ # --- UPDATED IMPORTS (LANGCHAIN v1.1 COMPATIBLE) ---
10
+ from langchain_core.documents import Document
11
+
12
+ from langchain_text_splitters import RecursiveCharacterTextSplitter
13
+ from langchain_community.vectorstores import FAISS
14
+ from langchain_core.prompts import ChatPromptTemplate
15
+ from langchain_huggingface import HuggingFaceEndpoint, HuggingFaceEndpointEmbeddings, ChatHuggingFace
16
+ # ... rest of your code ...
17
+ @st.cache_data
18
+ def get_pdf_text(pdf_docs):
19
+ """
20
+ Extracts text from PDFs and creates Document objects with metadata.
21
+ Returns a list of Document objects, not a string.
22
+ """
23
+ documents = []
24
+ for pdf in pdf_docs:
25
+ pdf_reader = PdfReader(pdf)
26
+ pdf_filename = pdf.name
27
+
28
+ for i, page in enumerate(pdf_reader.pages):
29
+ text = page.extract_text()
30
+ if text:
31
+ # Create a Document with metadata (page number starts at 1)
32
+ doc = Document(
33
+ page_content=text,
34
+ metadata={"page": i + 1, "source": pdf_filename}
35
+ )
36
+ documents.append(doc)
37
+ return documents
38
+
39
+ @st.cache_data
40
+ def get_text_chunks(documents):
41
+ """
42
+ Splits Document objects into chunks while preserving metadata.
43
+ """
44
+ text_splitter = RecursiveCharacterTextSplitter(
45
+ chunk_size=2000, # 10k is often too big for precise page tracking; 2k is better
46
+ chunk_overlap=200
47
+ )
48
+ # split_documents ensures metadata (page num) is copied to new chunks
49
+ chunks = text_splitter.split_documents(documents)
50
+ return chunks
51
+
52
+ # --- Hugging Face Model Functions ---
53
+ @st.cache_resource
54
+ def get_hf_embeddings():
55
+ hf_token = os.getenv("HF_API_TOKEN") or st.secrets.get("HF_API_TOKEN")
56
+ if not hf_token:
57
+ st.error("HF_API_TOKEN not found in environment variables or secrets.toml.")
58
+ return None
59
+ return HuggingFaceEndpointEmbeddings(
60
+ model="BAAI/bge-small-en-v1.5",
61
+ huggingfacehub_api_token=hf_token
62
+ )
63
+
64
+ @st.cache_resource
65
+ def get_hf_llm():
66
+ hf_token = os.getenv("HF_API_TOKEN") or st.secrets.get("HF_API_TOKEN")
67
+ if not hf_token:
68
+ return None
69
+
70
+ llm = HuggingFaceEndpoint(
71
+ huggingfacehub_api_token=hf_token,
72
+ repo_id="mistralai/Mistral-7B-Instruct-v0.2",
73
+ # repo_id="meta-llama/Llama-3.2-3B-Instruct",
74
+ temperature=0.3,
75
+ max_new_tokens=512,
76
+ task="conversational"
77
+ )
78
+ return ChatHuggingFace(llm=llm)
79
+
80
+ # --- Vector Store Functions (UPDATED) ---
81
+ def get_vector_store(text_chunks):
82
+ """Creates FAISS vector store from Document objects."""
83
+ if not text_chunks:
84
+ st.error("No text chunks found.")
85
+ return
86
+
87
+ embeddings = get_hf_embeddings()
88
+ if embeddings is None:
89
+ return
90
+
91
+ try:
92
+ # Use from_documents instead of from_texts to keep metadata
93
+ vector_store = FAISS.from_documents(documents=text_chunks, embedding=embeddings)
94
+ vector_store.save_local("faiss_index")
95
+ st.success("Vector store created and saved locally.")
96
+ except Exception as e:
97
+ st.error(f"Error creating vector store: {e}")
98
+
99
+ def load_faiss_index():
100
+ try:
101
+ embeddings = get_hf_embeddings()
102
+ if embeddings is None:
103
+ return None
104
+
105
+ db = FAISS.load_local(
106
+ "faiss_index",
107
+ embeddings,
108
+ allow_dangerous_deserialization=True
109
+ )
110
+ return db
111
+ except Exception as e:
112
+ # Silent fail on startup if index doesn't exist yet
113
+ return None
114
+
115
+ # --- Conversational Chain ---
116
+ def get_conversational_chain():
117
+ model = get_hf_llm()
118
+ if model is None:
119
+ return None
120
+
121
+ system_prompt = """
122
+ You are a helpful AI assistant answering questions based on PDF documents.
123
+ If the answer is not in the context, say you don't know and do not give any text about the question, just say "I don't know."
124
+ Context:
125
+ {context}
126
+ **Note:** please do not answer questions that are not related to the context provided. If the context does not contain the answer, respond with "I don't know."
127
+
128
+ """
129
+
130
+ prompt = ChatPromptTemplate.from_messages([
131
+ ("system", system_prompt),
132
+ # The user's question will be passed here as the "human" message
133
+ ("human", "Question: {input}")
134
+ ])
135
+
136
+ # Create a simple chain using the pipe operator (|)
137
+ chain = prompt | model
138
+ return chain
139
+
140
+ # --- User Input Handler (UPDATED) ---
141
+ def user_input(user_question, db):
142
+ if db is None:
143
+ st.warning("Vector store is not loaded.")
144
+ return
145
+
146
+ try:
147
+ # 1. Get relevant documents
148
+ docs = db.similarity_search(user_question, k=4)
149
+
150
+ # 2. Get the response
151
+ chain = get_conversational_chain()
152
+ if chain is None:
153
+ st.error("Check API Key")
154
+ return
155
+
156
+ response = chain.invoke({
157
+ "context": docs,
158
+ "input": user_question
159
+ })
160
+
161
+ # 3. Display Response
162
+ st.markdown("### Reply:")
163
+ # Extract just the content text from the response
164
+ if hasattr(response, 'content'):
165
+ # Response is an AIMessage object
166
+ st.write(response.content)
167
+ elif isinstance(response, dict) and "answer" in response:
168
+ st.write(response["answer"])
169
+ elif isinstance(response, dict) and "content" in response:
170
+ st.write(response["content"])
171
+ else:
172
+ # Try to get string representation
173
+ st.write(str(response))
174
+
175
+ # # 4. Display Page Numbers (New Feature)
176
+ # st.markdown("---")
177
+ # st.markdown("#### 📄 Source Pages:")
178
+
179
+ # # Use a set to avoid duplicate page numbers
180
+ # unique_sources = set()
181
+ # for doc in docs:
182
+ # page_num = doc.metadata.get("page", "Unknown")
183
+ # file_name = doc.metadata.get("source", "Unknown")
184
+ # unique_sources.add(f"{file_name} (Page {page_num})")
185
+
186
+ # # Display sources cleanly
187
+ # for source in unique_sources:
188
+ # st.markdown(f"- {source}")
189
+
190
+ except Exception as e:
191
+ st.error(f"Error: {e}")
192
+
193
+ # --- Main App ---
194
+ def main():
195
+ st.set_page_config(page_title="Chat PDF (Hugging Face)")
196
+ st.header("Chat with PDF using mistralai/Mistral-7B-Instruct-v0.2")
197
+
198
+ if "faiss_db" not in st.session_state:
199
+ st.session_state.faiss_db = None
200
+
201
+ # Attempt to load index on launch
202
+ if st.session_state.faiss_db is None and os.path.exists("faiss_index"):
203
+ st.session_state.faiss_db = load_faiss_index()
204
+
205
+ with st.sidebar:
206
+ st.title("Menu:")
207
+ pdf_docs = st.file_uploader("Upload PDFs", accept_multiple_files=True)
208
+
209
+ if st.button("Process PDFs"):
210
+ if pdf_docs:
211
+ with st.spinner("Processing..."):
212
+ # 1. Get Documents (not string)
213
+ raw_documents = get_pdf_text(pdf_docs)
214
+ # 2. Split Documents (preserves metadata)
215
+ text_chunks = get_text_chunks(raw_documents)
216
+ # 3. Create Vector Store
217
+ get_vector_store(text_chunks)
218
+
219
+ st.session_state.faiss_db = load_faiss_index()
220
+ else:
221
+ st.warning("Upload a PDF first.")
222
+
223
+ user_question = st.text_input("Ask a Question")
224
+
225
+ if user_question:
226
+ if st.session_state.faiss_db:
227
+ user_input(user_question, st.session_state.faiss_db)
228
+ else:
229
+ st.warning("Process PDFs first.")
230
+
231
+ if __name__ == "__main__":
232
+ main()