Spaces:
Sleeping
Sleeping
Upload app.py
Browse files
app.py
ADDED
@@ -0,0 +1,325 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import gradio as gr
|
3 |
+
from langchain.text_splitter import RecursiveCharacterTextSplitter
|
4 |
+
from langchain_community.document_loaders import PyPDFLoader, WebBaseLoader
|
5 |
+
from langchain_community.tools.tavily_search import TavilySearchResults
|
6 |
+
from langchain_community.vectorstores import SKLearnVectorStore
|
7 |
+
from langchain_openai import ChatOpenAI
|
8 |
+
from langchain_huggingface import HuggingFaceEmbeddings
|
9 |
+
from langchain_pinecone import PineconeVectorStore
|
10 |
+
from langchain.prompts import PromptTemplate
|
11 |
+
from langchain_core.output_parsers import StrOutputParser
|
12 |
+
from langchain_core.prompts import ChatPromptTemplate
|
13 |
+
from pydantic import BaseModel, Field
|
14 |
+
from typing import List, TypedDict, Optional
|
15 |
+
from langchain.schema import Document
|
16 |
+
from langgraph.graph import START, END, StateGraph
|
17 |
+
from dotenv import load_dotenv
|
18 |
+
|
19 |
+
load_dotenv()
|
20 |
+
|
21 |
+
url = [
|
22 |
+
"https://www.investopedia.com/",
|
23 |
+
"https://www.fool.com/",
|
24 |
+
"https://www.morningstar.com/",
|
25 |
+
"https://www.kiplinger.com/",
|
26 |
+
"https://www.nerdwallet.com/"
|
27 |
+
]
|
28 |
+
|
29 |
+
# Initialize Embedding and Vector DB
|
30 |
+
embedding_model = HuggingFaceEmbeddings(model_name="sentence-transformers/all-MiniLM-L6-v2")
|
31 |
+
|
32 |
+
# Initialize Pinecone connection
|
33 |
+
try:
|
34 |
+
pc = PineconeVectorStore(
|
35 |
+
pinecone_api_key=os.environ.get('PINECONE_KEY'),
|
36 |
+
embedding=embedding_model,
|
37 |
+
index_name='rag-rubic',
|
38 |
+
namespace='vectors_lightmodel'
|
39 |
+
)
|
40 |
+
retriever = pc.as_retriever(search_kwargs={"k": 10})
|
41 |
+
except Exception as e:
|
42 |
+
print(f"Pinecone connection error: {e}")
|
43 |
+
# Fallback to SKLearn vector store if Pinecone fails
|
44 |
+
retriever = None
|
45 |
+
|
46 |
+
# Initialize the LLM
|
47 |
+
llm = ChatOpenAI(
|
48 |
+
model='gpt-4o-mini',
|
49 |
+
api_key=os.environ.get('OPENAI_KEY'),
|
50 |
+
temperature=0.2
|
51 |
+
)
|
52 |
+
|
53 |
+
# Schema for grading documents
|
54 |
+
class GradeDocuments(BaseModel):
|
55 |
+
binary_score: str = Field(description="Documents are relevant to the question, 'yes' or 'no'")
|
56 |
+
|
57 |
+
structured_llm_grader = llm.with_structured_output(GradeDocuments)
|
58 |
+
|
59 |
+
# Define System and Grading prompt
|
60 |
+
system = """You are a grader assessing relevance of a retrieved document to a user question.
|
61 |
+
If the document contains keyword(s) or semantic meaning related to the question, grade it as relevant.
|
62 |
+
Give a binary score 'yes' or 'no' score to indicate whether the document is relevant to the question."""
|
63 |
+
|
64 |
+
grade_prompt = ChatPromptTemplate.from_messages([
|
65 |
+
("system", system),
|
66 |
+
("human", "Retrieved document: \n\n {documents} \n\n User question: {question}")
|
67 |
+
])
|
68 |
+
|
69 |
+
retrieval_grader = grade_prompt | structured_llm_grader
|
70 |
+
|
71 |
+
# RAG Prompt template
|
72 |
+
prompt = PromptTemplate(
|
73 |
+
template='''
|
74 |
+
You are a Registered Investment Advisor with expertise in Indian financial markets and client relations.
|
75 |
+
You must understand what the user is asking about their financial investments and respond to their queries based on the information in the documents only.
|
76 |
+
|
77 |
+
Use the following documents to answer the question. If you do not know the answer, say you don't know.
|
78 |
+
|
79 |
+
Query: {question}
|
80 |
+
Documents: {context}
|
81 |
+
''',
|
82 |
+
input_variables=['question', 'context']
|
83 |
+
)
|
84 |
+
|
85 |
+
rag_chain = prompt | llm | StrOutputParser()
|
86 |
+
|
87 |
+
# Web search tool for adding data from websites
|
88 |
+
web_search_tool = TavilySearchResults(api_key=os.environ.get('TAVILY_API_KEY'), k=5)
|
89 |
+
|
90 |
+
# Load website data
|
91 |
+
try:
|
92 |
+
print("Loading web data...")
|
93 |
+
docs = []
|
94 |
+
for i in url:
|
95 |
+
try:
|
96 |
+
docs.append(WebBaseLoader(i).load())
|
97 |
+
except Exception as e:
|
98 |
+
print(f"Error loading {i}: {e}")
|
99 |
+
|
100 |
+
docs_list = [item for sublist in docs for item in sublist]
|
101 |
+
|
102 |
+
# Split documents into chunks
|
103 |
+
text_splitter = RecursiveCharacterTextSplitter.from_tiktoken_encoder(
|
104 |
+
chunk_size=1000,
|
105 |
+
chunk_overlap=100
|
106 |
+
)
|
107 |
+
doc_splits = text_splitter.split_documents(docs_list)
|
108 |
+
|
109 |
+
# VectorStore from the web-scraped documents
|
110 |
+
vectorstore = SKLearnVectorStore.from_documents(
|
111 |
+
documents=doc_splits,
|
112 |
+
embedding=embedding_model
|
113 |
+
)
|
114 |
+
retriever_web = vectorstore.as_retriever(search_kwargs={"k": 5})
|
115 |
+
print(f"Loaded {len(doc_splits)} document chunks from web sources")
|
116 |
+
except Exception as e:
|
117 |
+
print(f"Error in web data processing: {e}")
|
118 |
+
# Create a simple retriever that returns empty results if web loading fails
|
119 |
+
retriever_web = lambda x: []
|
120 |
+
|
121 |
+
# Define Graph states and transitions
|
122 |
+
class GraphState(TypedDict):
|
123 |
+
question: str
|
124 |
+
generation: Optional[str]
|
125 |
+
need_web_search: Optional[str] # Changed from 'web_search' to 'need_web_search'
|
126 |
+
documents: List
|
127 |
+
|
128 |
+
def retrieve_db(state):
|
129 |
+
"""Gather data for the query."""
|
130 |
+
question = state['question']
|
131 |
+
if retriever:
|
132 |
+
try:
|
133 |
+
results = retriever.invoke(question)
|
134 |
+
return {'documents': results, 'question': question}
|
135 |
+
except Exception as e:
|
136 |
+
print(f"Retriever error: {e}")
|
137 |
+
|
138 |
+
# If retriever fails or doesn't exist, return empty documents
|
139 |
+
return {'documents': [], 'question': question, 'need_web_search': 'yes'}
|
140 |
+
|
141 |
+
def grade_docs(state):
|
142 |
+
"""Grades the docs generated by the retriever_db"""
|
143 |
+
question = state['question']
|
144 |
+
docs = state['documents']
|
145 |
+
|
146 |
+
if not docs:
|
147 |
+
return {"documents": [], 'question': question, 'need_web_search': 'yes'}
|
148 |
+
|
149 |
+
filtered_data = []
|
150 |
+
web_search_needed = "no"
|
151 |
+
|
152 |
+
try:
|
153 |
+
for doc in docs:
|
154 |
+
doc_content = doc.page_content if hasattr(doc, 'page_content') else str(doc)
|
155 |
+
score = retrieval_grader.invoke({'question': question, 'documents': doc_content})
|
156 |
+
grade = score.binary_score
|
157 |
+
if grade == 'yes':
|
158 |
+
filtered_data.append(doc)
|
159 |
+
except Exception as e:
|
160 |
+
print(f"Error in document grading: {e}")
|
161 |
+
web_search_needed = "yes"
|
162 |
+
|
163 |
+
# If no relevant documents were found, trigger web search
|
164 |
+
if not filtered_data:
|
165 |
+
web_search_needed = "yes"
|
166 |
+
|
167 |
+
return {
|
168 |
+
"documents": filtered_data,
|
169 |
+
'question': question,
|
170 |
+
'need_web_search': web_search_needed # Updated key name
|
171 |
+
}
|
172 |
+
|
173 |
+
def decide(state):
|
174 |
+
"""Decide if the generation should be based on DB or web search DATA"""
|
175 |
+
web = state.get('need_web_search', 'no') # Updated key name
|
176 |
+
if web == 'yes':
|
177 |
+
return 'web_search'
|
178 |
+
else:
|
179 |
+
return 'generate'
|
180 |
+
|
181 |
+
def web_search(state):
|
182 |
+
"""Based on the Grade, will proceed with WebSearch within the given URL's."""
|
183 |
+
question = state['question']
|
184 |
+
documents = state.get("documents", [])
|
185 |
+
|
186 |
+
try:
|
187 |
+
# First try website-specific retriever
|
188 |
+
docs = retriever_web.invoke(question)
|
189 |
+
if not docs:
|
190 |
+
# If no results, try Tavily search
|
191 |
+
search_results = web_search_tool.invoke(question)
|
192 |
+
data = "\n".join(result["content"] for result in search_results)
|
193 |
+
docs = [Document(page_content=data)]
|
194 |
+
except Exception as e:
|
195 |
+
print(f"Web search error: {e}")
|
196 |
+
# Create a fallback document if search fails
|
197 |
+
docs = [Document(page_content="Unable to retrieve additional information.")]
|
198 |
+
|
199 |
+
# Combine existing documents with new ones
|
200 |
+
all_docs = documents + docs
|
201 |
+
|
202 |
+
return {'documents': all_docs, 'question': question}
|
203 |
+
|
204 |
+
def generate(state):
|
205 |
+
"""Generate response based on retrieved documents"""
|
206 |
+
documents = state.get('documents', [])
|
207 |
+
question = state['question']
|
208 |
+
|
209 |
+
# Convert documents to text for the context
|
210 |
+
if documents:
|
211 |
+
try:
|
212 |
+
context = "\n\n".join(
|
213 |
+
doc.page_content if hasattr(doc, 'page_content') else str(doc)
|
214 |
+
for doc in documents
|
215 |
+
)
|
216 |
+
except Exception as e:
|
217 |
+
print(f"Error processing documents: {e}")
|
218 |
+
context = "Error retrieving relevant information."
|
219 |
+
else:
|
220 |
+
context = "No relevant information found."
|
221 |
+
|
222 |
+
try:
|
223 |
+
response = rag_chain.invoke({'context': context, 'question': question})
|
224 |
+
except Exception as e:
|
225 |
+
print(f"Generation error: {e}")
|
226 |
+
response = "I apologize, but I encountered an error while generating a response. Please try asking your question again."
|
227 |
+
|
228 |
+
return {
|
229 |
+
'documents': documents,
|
230 |
+
'question': question,
|
231 |
+
'generation': response
|
232 |
+
}
|
233 |
+
|
234 |
+
# Compile Workflow
|
235 |
+
workflow = StateGraph(GraphState)
|
236 |
+
workflow.add_node("retrieve", retrieve_db)
|
237 |
+
workflow.add_node("grader", grade_docs)
|
238 |
+
workflow.add_node("web_search", web_search) # Now this won't conflict with the state key
|
239 |
+
workflow.add_node("generate", generate)
|
240 |
+
|
241 |
+
workflow.add_edge(START, "retrieve")
|
242 |
+
workflow.add_edge("retrieve", "grader")
|
243 |
+
workflow.add_conditional_edges(
|
244 |
+
"grader",
|
245 |
+
decide,
|
246 |
+
{
|
247 |
+
'web_search': 'web_search',
|
248 |
+
'generate': 'generate'
|
249 |
+
},
|
250 |
+
)
|
251 |
+
workflow.add_edge("web_search", "generate")
|
252 |
+
workflow.add_edge("generate", END)
|
253 |
+
|
254 |
+
# Compile the graph
|
255 |
+
crag = workflow.compile()
|
256 |
+
|
257 |
+
# Define Gradio Interface with proper chat history management
|
258 |
+
def process_query(user_input, history):
|
259 |
+
# Initialize history if it's None
|
260 |
+
if history is None:
|
261 |
+
history = []
|
262 |
+
|
263 |
+
# Add user input to history
|
264 |
+
history.append((user_input, ""))
|
265 |
+
|
266 |
+
# Process the query
|
267 |
+
inputs = {"question": user_input}
|
268 |
+
response = ""
|
269 |
+
|
270 |
+
try:
|
271 |
+
# Execute the graph
|
272 |
+
result = crag.invoke(inputs)
|
273 |
+
if result and 'generation' in result:
|
274 |
+
response = result['generation']
|
275 |
+
else:
|
276 |
+
response = "I couldn't find relevant information to answer your question."
|
277 |
+
except Exception as e:
|
278 |
+
print(f"Error in crag execution: {e}")
|
279 |
+
response = "I encountered an error while processing your request. Please try again."
|
280 |
+
|
281 |
+
# Update the last response in history
|
282 |
+
history[-1] = (user_input, response)
|
283 |
+
|
284 |
+
return history, ""
|
285 |
+
|
286 |
+
# Gradio Interface
|
287 |
+
with gr.Blocks() as demo:
|
288 |
+
gr.Markdown("# 🤖 RAG-Powered Financial Advisor Chatbot")
|
289 |
+
|
290 |
+
chatbot = gr.Chatbot(
|
291 |
+
[],
|
292 |
+
elem_id="chatbot",
|
293 |
+
bubble_full_width=False,
|
294 |
+
height=600,
|
295 |
+
avatar_images=(None, "🤖")
|
296 |
+
)
|
297 |
+
|
298 |
+
with gr.Row():
|
299 |
+
msg = gr.Textbox(
|
300 |
+
placeholder="Ask me anything about Indian financial markets...",
|
301 |
+
label="Your question:",
|
302 |
+
scale=9
|
303 |
+
)
|
304 |
+
submit_btn = gr.Button("Send", scale=1)
|
305 |
+
|
306 |
+
clear_btn = gr.Button("Clear Chat")
|
307 |
+
|
308 |
+
# Set up event handlers
|
309 |
+
submit_click_event = submit_btn.click(
|
310 |
+
process_query,
|
311 |
+
inputs=[msg, chatbot],
|
312 |
+
outputs=[chatbot, msg]
|
313 |
+
)
|
314 |
+
|
315 |
+
msg.submit(
|
316 |
+
process_query,
|
317 |
+
inputs=[msg, chatbot],
|
318 |
+
outputs=[chatbot, msg]
|
319 |
+
)
|
320 |
+
|
321 |
+
clear_btn.click(lambda: [], outputs=[chatbot])
|
322 |
+
|
323 |
+
|
324 |
+
if __name__ == "__main__":
|
325 |
+
demo.launch()
|