Haspit / app.py
nitind1's picture
Upload app.py with huggingface_hub
a6ea608 verified
Raw History Blame Contribute Delete
25.4 kB
######################## WRITE YOUR CODE HERE #########################
# Import libraries
import os
import chromadb
from dotenv import load_dotenv
import json
# LangChain imports
from langchain_core.documents import Document
from langchain_core.runnables import RunnablePassthrough
from langchain_core.output_parsers import StrOutputParser
from langchain.prompts import ChatPromptTemplate
from langchain.chains.query_constructor.base import AttributeInfo
from langchain.retrievers.self_query.base import SelfQueryRetriever
from langchain.retrievers.document_compressors import LLMChainExtractor, CrossEncoderReranker
from langchain.retrievers import ContextualCompressionRetriever
# LangChain community & experimental imports
from langchain_community.vectorstores import Chroma
from langchain_community.document_loaders import PyPDFDirectoryLoader, PyPDFLoader
from langchain_community.cross_encoders import HuggingFaceCrossEncoder
from langchain_experimental.text_splitter import SemanticChunker
from langchain.text_splitter import (
CharacterTextSplitter,
RecursiveCharacterTextSplitter
)
from langchain_core.tools import tool
from langchain.agents import create_tool_calling_agent, AgentExecutor
from langchain_core.prompts import ChatPromptTemplate
# LangChain OpenAI imports
from langchain_openai import AzureOpenAIEmbeddings, AzureChatOpenAI
from langchain.memory import ConversationSummaryBufferMemory
from langchain_openai import ChatOpenAI
from langchain_community.embeddings import OpenAIEmbeddings
from langchain_openai import OpenAIEmbeddings
from langchain_community.embeddings.sentence_transformer import SentenceTransformerEmbeddings
from langchain_community.vectorstores import Chroma
from langchain_community.utilities.sql_database import SQLDatabase
from langchain_community.agent_toolkits import create_sql_agent
from langchain_core.prompts import ChatPromptTemplate, MessagesPlaceholder
from langchain_core.tools import tool
from langchain.agents import create_tool_calling_agent, AgentExecutor
# LlamaParse & LlamaIndex imports
from llama_parse import LlamaParse
from llama_index.core import Settings, SimpleDirectoryReader
# LangGraph import
from langgraph.graph import StateGraph, END, START
# Pydantic import
from pydantic import BaseModel
# Typing imports
from typing import Dict, List, Tuple, Any, TypedDict
# Other utilities
import numpy as np
from groq import Groq
from mem0 import MemoryClient
import streamlit as st
from datetime import datetime
#====================================SETUP=====================================#
# Fetch secrets from Hugging Face Spaces
api_key = os.environ["API_KEY"]
endpoint = os.environ["OPENAI_API_BASE"]
llama_api_key = os.environ['LLAMA_API_KEY']
#mem0_api_key = os.environ['mem0']
# Initializing the OpenAI embedding function for Chroma
embedding_function = chromadb.utils.embedding_functions.OpenAIEmbeddingFunction(
api_base=endpoint, # Complete the code to define the API base endpoint
api_key=api_key, # Complete the code to define the API key
model_name='text-embedding-ada-002' # This is a fixed value and does not need modification
)
# Initializing the OpenAI Embeddings
embedding_model = OpenAIEmbeddings(
openai_api_base=endpoint,
openai_api_key=api_key,
model='text-embedding-ada-002'
)
# Initializing the Chat OpenAI model
llm = ChatOpenAI(
openai_api_base=endpoint,
openai_api_key=api_key,
model="gpt-4o-mini",
streaming=False
)
# set the LLM and embedding model in the LlamaIndex settings.
Settings.llm = llm # Complete the code to define the LLM model
Settings.embedding = embedding_model # Complete the code to define the embedding model
#================================Creating Langgraph agent======================#
class AgentState(TypedDict):
query: str # The current user query
expanded_query: str # The expanded version of the user query
context: List[Dict[str, Any]] # Retrieved documents (content and metadata)
response: str # The generated response to the user query
precision_score: float # The precision score of the response
#groundedness_score: float # The groundedness score of the response
#groundedness_loop_count: int # Counter for groundedness refinement loops
precision_loop_count: int # Counter for precision refinement loops
feedback: str
query_feedback: str
#groundedness_check: bool
loop_max_iter: int
def expand_query(state: AgentState) -> AgentState:
"""
Expands the user query to improve retrieval of information using few-shot prompting.
Args:
state (Dict): The current state of the workflow, containing the user query.
Returns:
Dict: The updated state with the expanded query.
"""
system_message = '''Perform query expansion on the query received. If there are multiple common ways of phrasing a user query \
or common synonyms for key words in the query, make sure to return multiple versions \
of the query with the different phrasings.
If the query has multiple parts, split them into separate simpler queries. This is the only case where you can generate more than 3 queries.
If there are acronyms or words you are not familiar with, do not try to rephrase them.
Return only 3 versions of the query as a list.
Generate only a list of queries. Do not mention anything before or after the list.
Question:
{query}'''
expand_prompt = ChatPromptTemplate.from_messages([
("system", system_message),
("user", "{query}")
])
chain = expand_prompt | llm | StrOutputParser()
state['expanded_query'] = chain.invoke({"query": state['query']})
return state
# Initialize the Chroma vector store for retrieving documents
vectorstore = Chroma(
collection_name="haspit1",
persist_directory="./haspit1_db",
embedding_function=embedding_model
)
# Create a retriever from the vector store
retriever = vectorstore.as_retriever(
search_type='similarity',
search_kwargs={'k': 5}
)
def retrieve_documents(query: str):
retrieved_documents = []
retrieved_documents = retriever.invoke("Fetch at most 5 documents to answer"+str(query))
retrieved_documents.extend(retriever.invoke("Fetch at most 5 tables to answer "+ str(query)))
return retrieved_documents
def retrieve_context(state: AgentState) -> AgentState:
"""
Retrieves documents from the vector store using the expanded or original query.
Args:
state (Dict): The current state of the workflow, containing the query and expanded query.
Returns:
Dict: The updated state with the retrieved documents.
"""
query = state['expanded_query']
#print("Query used for retrieval:", query) # Debugging: Print the query
# Retrieve documents from the vector store
#docs = retriever.invoke(query)
context_list = []
#for query in state['expanded_query']:
#print("Query used for retrieval:", query) # Debugging: Print the query
retrieved = retrieve_documents(query)
#print("Retrieved documents:", retrieved)
#docs = retriever.invoke(query)
context_list.extend(retrieved)
#print(context_list)
#print("Retrieved documents:", retrieved)
#return only unique docs
unique_context_dict = {doc.metadata['source'].split("/")[-1]: doc for doc in context_list}
unique_context_list = list(unique_context_dict.values())
#print(unique_context_list)
updated_state = state.copy() # Create a shallow copy to avoid modifying the original
updated_state['context'] = unique_context_list
#state['context'] = unique_context_list
#print("Unique documents retrieved")
#print("Unique documents retrieved", unique_context_list)
#return unique_context_list
#print("Retrieved documents:", docs) # Debugging: Print the raw docs object
# Extract both page_content and metadata from each document
#print("Document: ", doc.metadata['source'].split("/")[-1])
#context= [
#{
#"Document": doc.metadata['source'].split("/")[-1],
#"page": doc.metadata['page'], # The page number or other metadata
#"content": doc.page_content # The actual content of the document
#"metadata": doc.metadata # The metadata (e.g., source, page number, etc.)
#}
#for doc in docs
#]
#state['context'] = context
#print(updated_state['context']) # Debugging: Print the extracted context
return updated_state
def check_precision(state: Dict) -> Dict:
"""
Checks whether the context precisely addresses the user’s query.
Args:
state (Dict): The current state of the workflow, containing the query and response.
Returns:
Dict: The updated state with the precision score.
"""
#print("---------check_precision---------")
system_message = '''Given the query and context verify if the context was useful in arriving at the answer.
Give precision score of "1" if useful and "0" if it was not useful.
'''
precision_prompt = ChatPromptTemplate.from_messages([
("system", system_message),
("user", "Query: {query}\nResponse: {context}\n\nPrecision score:")
])
chain = precision_prompt | llm | StrOutputParser()
precision_score = float(chain.invoke({
"query": state['query'],
"context": state['context']
}))
state['precision_score'] = precision_score
#print("precision_score:", precision_score)
state['precision_loop_count'] += 1
#print("#########Precision Incremented###########")
return state
def refine_query(state: Dict) -> Dict:
"""
Suggests improvements for the expanded query.
Args:
state (Dict): The current state of the workflow, containing the query and expanded query.
Returns:
Dict: The updated state with query refinement suggestions.
"""
#print("---------refine_query---------")
system_message = '''Your task is to refine the expanded query to improve the precision of the context retrieved based on the feedback provided.'''
refine_query_prompt = ChatPromptTemplate.from_messages([
("system", system_message),
("user", "Original Query: {query}\nExpanded Query: {expanded_query}\n\n"
"What improvements can be made for a better search?")
])
chain = refine_query_prompt | llm | StrOutputParser()
# Store refinement suggestions without modifying the original expanded query
query_feedback = f"Previous Expanded Query: {state['expanded_query']}\nSuggestions: {chain.invoke({'query': state['query'], 'expanded_query': state['expanded_query']})}"
#print("query_feedback: ", query_feedback)
#print(f"Groundedness loop count: {state['groundedness_loop_count']}")
state['query_feedback'] = query_feedback
return state
def should_continue_precision(state: Dict) -> str:
"""Decides if precision is sufficient or needs improvement."""
#print("---------should_continue_precision---------")
#print("precision loop count: ", state['precision_loop_count'])
if state['precision_score']>=1: # Threshold for precision
return "pass" # Complete the workflow
else:
if state['precision_loop_count']> state['loop_max_iter']: # Maximum allowed loops
return "max_iterations_reached"
else:
#print(f"---------Precision Score Threshold Not met. Refining Query-----------") # Debugging
return "refine_query" # Refine the query
def max_iterations_reached(state: AgentState) -> AgentState:
"""Handles the case where max iterations are reached."""
state['response'] = "We need additional information to search for relevant documents."
return state
from langgraph.graph import END, StateGraph, START
def create_workflow() -> StateGraph:
"""Creates the updated workflow for the AI agent."""
workflow = StateGraph(AgentState)
# Add processing nodes
workflow.add_node("expand_query", expand_query) # Step 1: Expand user query.
workflow.add_node("retrieve_context", retrieve_context) # Step 2: Retrieve relevant documents.
#workflow.add_node("craft_response", craft_response) # Step 3: Generate a response based on retrieved data.
#workflow.add_node("score_groundedness", score_groundedness) # Step 4: Evaluate response grounding.
#workflow.add_node("refine_response", refine_response) # Step 5: Improve response if it's weakly grounded.
workflow.add_node("check_precision", check_precision) # Step 6: Evaluate response precision.
workflow.add_node("refine_query",refine_query ) # Step 7: Improve query if response lacks precision. Complete with the function to refine the query
workflow.add_node("max_iterations_reached", max_iterations_reached) # Step 8: Handle max iterations.
# Main flow edges
workflow.add_edge(START, "expand_query")
workflow.add_edge("expand_query", "retrieve_context")
workflow.add_edge("retrieve_context", "check_precision")
#workflow.add_edge("craft_response", "score_groundedness")
# Conditional edges based on groundedness check
#workflow.add_conditional_edges(
#"score_groundedness",
#should_continue_groundedness, # Use the conditional function
#{
#"check_precision": "check_precision", # If well-grounded, proceed to precision check.
#"refine_response": "refine_response", # If not, refine the response.
#"max_iterations_reached": "max_iterations_reached" # If max loops reached, exit.
#}
#)
#workflow.add_edge("refine_response", "craft_response") # Refined responses are reprocessed.
# Conditional edges based on precision check
workflow.add_conditional_edges(
"check_precision",
should_continue_precision, # Use the conditional function
{
"pass": END, # If precise, complete the workflow.
"refine_query": "refine_query", # If imprecise, refine the query.
"max_iterations_reached": "max_iterations_reached" # If max loops reached, exit.
}
)
workflow.add_edge("refine_query", "expand_query") # Refined queries go through expansion again.
workflow.add_edge("max_iterations_reached", END)
return workflow
#=========================== Defining the agentic rag tool ====================#
WORKFLOW_APP = create_workflow().compile()
@tool
def agentic_rag(query: str):
"""
Runs the RAG-based agent with conversation history for context-aware responses.
Args:
query (str): The current user query.
Returns:
Dict: The retireved response.
"""
# Initialize state with necessary parameters
inputs = {
"query": query,
"expanded_query": "",
"context": [],
"response": "",
"precision_score": 0,
#"groundedness_score":0,
#"groundedness_loop_count": 5,
"precision_loop_count": 0,
"feedback": "",
"query_feedback": "",
"loop_max_iter": 5
}
output = WORKFLOW_APP.invoke(inputs)
return output['context']
#================================ Guardrails ===========================#
llama_guard_client = Groq(api_key=llama_api_key)
# Function to filter user input with Llama Guard
def filter_input_with_llama_guard(user_input, model="llama-guard-3-8b"):
"""
Filters user input using Llama Guard to ensure it is safe.
Parameters:
- user_input: The input provided by the user.
- model: The Llama Guard model to be used for filtering (default is "llama-guard-3-8b").
Returns:
- The filtered and safe input.
"""
try:
# Create a request to Llama Guard to filter the user input
response = llama_guard_client.chat.completions.create(
messages=[{"role": "user", "content": user_input}],
model=model,
)
# Return the filtered input
return response.choices[0].message.content.strip()
except Exception as e:
print(f"Error with Llama Guard: {e}")
return None
#============================= Adding Memory to the agent using mem0 ===============================#
class HaspitBot:
def __init__(self):
"""
Initialize the SemanticBot class, setting up memory, the LLM client, tools, and the agent executor.
"""
# Initialize a memory client to store and retrieve customer interactions
#self.memory = MemoryClient(api_key=userdata.get("mem0"))
#self.memory = MemoryClient(userdata.get(api_key="mem0_api_key")) # Complete the code to define the memory client API key
# Initialize the OpenAI client using the provided credentials
self.client = ChatOpenAI(
model_name="gpt-4o-mini", # Specify the model to use (e.g., GPT-4 optimized version)
api_key= os.environ["API_KEY"], # API key for authentication
openai_api_base = os.environ["OPENAI_API_BASE"],
temperature=0 # Controls randomness in responses; 0 ensures deterministic results
)
# Define tools available to the chatbot, such as web search
tools = [agentic_rag]
# Define the system prompt to set the behavior of the chatbot
system_prompt = """
You are a knowledgeable agent, specializing in retrieving relevant documents.
Guidelines for Interaction:
Always use the agentic_rag tool to retrieve up-to-date documents.
"""
# Build the prompt template for the agent
prompt = ChatPromptTemplate.from_messages([
("system", system_prompt), # System instructions
("human", "{input}"), # Placeholder for human input
("placeholder", "{agent_scratchpad}") # Placeholder for intermediate reasoning steps
])
# Create an agent capable of interacting with tools and executing tasks
agent = create_tool_calling_agent(self.client, tools, prompt)
# Wrap the agent in an executor to manage tool interactions and execution flow
self.agent_executor = AgentExecutor(agent=agent, tools=tools, verbose=True)
#def store_customer_interaction(self, user_id: str, message: str, response: str, metadata: Dict = None):
#"""
#Store customer interaction in memory for future reference.
#Args:
#user_id (str): Unique identifier for the customer.
#message (str): Customer's query or message.
#response (str): Chatbot's response.
#metadata (Dict, optional): Additional metadata for the interaction.
#"""
#if metadata is None:
#metadata = {}
# Add a timestamp to the metadata for tracking purposes
#metadata["timestamp"] = datetime.now().isoformat()
# Format the conversation for storage
#conversation = [
#{"role": "user", "content": message},
#{"role": "assistant", "content": response}
#]
# Store the interaction in the memory client
#self.memory.add(
#conversation,
#user_id=user_id,
#output_format="v1.1",
#metadata=metadata
#)
#def get_relevant_history(self, user_id: str, query: str) -> List[Dict]:
#"""
#Retrieve past interactions relevant to the current query.
#Args:
#user_id (str): Unique identifier for the customer.
#query (str): The customer's current query.
#Returns:
#List[Dict]: A list of relevant past interactions.
#"""
#return self.memory.search(
#query=query, # Search for interactions related to the query
#user_id=user_id, # Restrict search to the specific user
#limit=5 # Complete the code to define the limit for retrieved interactions
#)
def handle_customer_query(self, query: str) -> str:
"""
Process a customer's query and provide a response, taking into account past interactions.
Args:
query (str): Customer's query.
Returns:
str: Chatbot's response.
"""
# Retrieve relevant past interactions for context
#relevant_history = self.get_relevant_history(user_id, query)
# Build a context string from the relevant history
#context = "Previous relevant interactions:\n"
#for memory in relevant_history:
#context += f"Customer: {memory['memory']}\n" # Customer's past messages
#context += f"Support: {memory['memory']}\n" # Chatbot's past responses
#context += "---\n"
# Print context for debugging purposes
#print("Context: ", context)
# Prepare a prompt combining past context and the current query
prompt = f"""
Current customer query: {query}
Retrieve relevant documents that match the customer query.
"""
# Generate a response using the agent
response = self.agent_executor.invoke({"input": prompt})
# Store the current interaction for future reference
#self.store_customer_interaction(
#user_id=user_id,
#message=query,
#response=response["output"],
#metadata={"type": "support_query"}
#)
# Return the chatbot's response
return response['output']
#=====================User Interface using streamlit ===========================#
def haspit_search_streamlit():
"""
A Streamlit-based UI for Semantic Search.
"""
st.title("Haspit AI Search Bot")
st.write("Ask me to search for relevant documents on Haspit.")
#st.write("Type 'exit' to end the conversation.")
# Initialize session state for chat history and user_id if they don't exist
if 'chat_history' not in st.session_state:
st.session_state.chat_history = []
if 'user_id' not in st.session_state:
st.session_state.user_id = None
# Login form: Only if user is not logged in
if st.session_state.user_id is None:
with st.form("login_form", clear_on_submit=True):
user_id = st.text_input("Please enter your name to begin:")
submit_button = st.form_submit_button("Login")
if submit_button and user_id:
st.session_state.user_id = user_id
st.session_state.chat_history.append({
"role": "assistant",
"content": f"Welcome, {user_id}! How can I help you with your search today?"
})
st.session_state.login_submitted = True # Set flag to trigger rerun
if st.session_state.get("login_submitted", False):
st.session_state.pop("login_submitted")
st.rerun()
else:
# Display chat history
for message in st.session_state.chat_history:
with st.chat_message(message["role"]):
st.write(message["content"])
# Chat input with custom placeholder text
user_query = st.chat_input("Type your search here (or exit to end): ", key="chat_input")
if user_query:
if user_query.lower() == "exit":
st.session_state.chat_history.append({"role": "user", "content": "exit"})
with st.chat_message("user"):
st.write("exit")
goodbye_msg = "Goodbye! Feel free to return if you want to search for content in documents on Haspit."
st.session_state.chat_history.append({"role": "assistant", "content": goodbye_msg})
with st.chat_message("assistant"):
st.write(goodbye_msg)
st.session_state.user_id = None
st.rerun()
return
st.session_state.chat_history.append({"role": "user", "content": user_query})
with st.chat_message("user"):
st.write(user_query)
# Filter input using Llama Guard
filtered_result = filter_input_with_llama_guard(user_query)
filtered_result = filtered_result.replace("\n", " ") # Normalize the result
# Check if input is safe based on allowed statuses
if filtered_result in ["safe", "unsafe S6", "unsafe S7"]:
try:
if 'chatbot' not in st.session_state:
st.session_state.chatbot = HaspitBot()
response = st.session_state.chatbot.handle_customer_query(user_query)
st.write(response)
st.session_state.chat_history.append({"role": "assistant", "content": response})
except Exception as e:
error_msg = f"Sorry, I encountered an error while processing your query. Please try again. Error: {str(e)}"
st.write(error_msg)
st.session_state.chat_history.append({"role": "assistant", "content": error_msg})
else:
inappropriate_msg = "I apologize, but I cannot process that input as it may be inappropriate. Please try again."
st.write(inappropriate_msg)
st.session_state.chat_history.append({"role": "assistant", "content": inappropriate_msg})
if __name__ == "__main__":
haspit_search_streamlit()