AI Notes

Building a Personalized Banking Chat Agent with React.js, RAG, LLM, and Redis with sample code

Here we outline a more detailed structure with conceptual sample code snippets for each layer of a conceptual personalized bank FAQ chat agent. Keep in mind that this is a simplified illustration, and a production-ready system would involve more robust error handling, security measures, and integration logic.

I. Knowledge Base Preparation:

Step 1: Data Collection & Structuring

Assume you have your bank’s FAQs in a structured format, perhaps JSON files where each entry has a question and an answer, or markdown files.

JSON

[
  {
    "question": "What are your current mortgage rates?",
    "answer": "Our current mortgage rates vary depending on the loan type and your credit score. Please visit our mortgage page or contact a loan officer for personalized rates."
  },
  {
    "question": "How do I reset my online banking password?",
    "answer": "To reset your online banking password, please click on the 'Forgot Password' link on the login page and follow the instructions."
  },
  // ... more FAQs
]

Step 2: Chunking

For larger documents (like policy documents), you’ll need to break them into smaller chunks. A simple approach is to split by paragraphs or sentences, ensuring context isn’t lost.

def chunk_text(text, chunk_size=512, overlap=50):
    chunks = []
    stride = chunk_size - overlap
    for i in range(0, len(text), stride):
        chunk = text[i:i + chunk_size]
        chunks.append(chunk)
    return chunks

# Example for a policy document
policy_text = """
This is a long banking policy document... It contains important information about accounts... and transaction limits...
Another paragraph discussing security measures... and fraud prevention...
"""
policy_chunks = chunk_text(policy_text)
print(f"Number of policy chunks: {len(policy_chunks)}")

Step 3: Embedding Generation

You’ll use an embedding model (e.g., from OpenAI, Sentence Transformers) to convert your FAQ answers and document chunks into vector embeddings.

Python

from sentence_transformers import SentenceTransformer
import numpy as np

embedding_model = SentenceTransformer('all-MiniLM-L6-v2')

faq_data = [
    {"question": "...", "answer": "Answer 1"},
    {"question": "...", "answer": "Answer 2"},
    # ...
]

faq_embeddings = embedding_model.encode([item["answer"] for item in faq_data])
print(f"Shape of FAQ embeddings: {faq_embeddings.shape}")

policy_chunks = ["chunk 1 of policy", "chunk 2 of policy"]
policy_embeddings = embedding_model.encode(policy_chunks)
print(f"Shape of policy embeddings: {policy_embeddings.shape}")

Step 4: Storing Embeddings in

You’ll use Redis with a vector search module (like Redis Stack) to store and index these embeddings.

Python

import redis
from redis.commands.search.field import TextField, VectorField
from redis.commands.search.indexDefinition import IndexDefinition, IndexType

REDIS_HOST = "localhost"
REDIS_PORT = 6379
REDIS_PASSWORD = None
INDEX_NAME = "bank_faq_embeddings"
VECTOR_DIM = 384  # Dimension of all-MiniLM-L6-v2 embeddings
NUM_VECTORS = len(faq_data) + len(policy_chunks)

r = redis.Redis(host=REDIS_HOST, port=REDIS_PORT, password=REDIS_PASSWORD)

# Define the schema for the Redis index
schema = (
    TextField("content"),  # Store the original text chunk
    VectorField("embedding", "FLAT", {"TYPE": "FLOAT32", "DIM": VECTOR_DIM, "DISTANCE_METRIC": "COSINE"})
)

# Define the index
definition = IndexDefinition(prefix=["faq:", "policy:"], index_type=IndexType.FLAT)

try:
    r.ft(INDEX_NAME).info()
    print(f"Index '{INDEX_NAME}' already exists.")
except:
    r.ft(INDEX_NAME).create_index(fields=schema, definition=definition)
    print(f"Index '{INDEX_NAME}' created.")

# Store FAQ embeddings
for i, item in enumerate(faq_data):
    key = f"faq:{i}"
    embedding = faq_embeddings[i].astype(np.float32).tobytes()
    r.hset(key, mapping={"content": item["answer"], "embedding": embedding})

# Store policy chunk embeddings
for i, chunk in enumerate(policy_chunks):
    key = f"policy:{i}"
    embedding = policy_embeddings[i].astype(np.float32).tobytes()
    r.hset(key, mapping={"content": chunk, "embedding": embedding})

print(f"Stored {r.ft(INDEX_NAME).info()['num_docs']} vectors in Redis.")

II. Implementation (Backend – Python/Node.js with a Framework like Flask/Express):

Python

from flask import Flask, request, jsonify
from sentence_transformers import SentenceTransformer
import redis
from redis.commands.search.query import Query

app = Flask(__name__)
embedding_model = SentenceTransformer('all-MiniLM-L6-v2')
r = redis.Redis(host=REDIS_HOST, port=REDIS_PORT, password=REDIS_PASSWORD)
INDEX_NAME = "bank_faq_embeddings"
VECTOR_DIM = 384
LLM_API_KEY = "YOUR_LLM_API_KEY" # Replace with your actual  key

def retrieve_relevant_documents(query, top_n=3):
    query_embedding = embedding_model.encode(query).astype(np.float32).tobytes()
    redis_query = (
        Query("*=>[KNN $topK @embedding $query_vector AS score]")
        .sort_by("score")
        .return_fields("content", "score")
        .dialect(2)
    )
    results = r.ft(INDEX_NAME).search(
        redis_query,
        query_params={"query_vector": query_embedding, "topK": top_n}
    )
    return [{"content": doc.content, "score": doc.score} for doc in results.docs]

def generate_response(query, context_documents):
    context = "\n".join([doc["content"] for doc in context_documents])
    prompt = f"""You are a helpful bank assistant. Use the following information to answer the user's question.
    If you cannot find the answer in the provided context, truthfully say "I'm sorry, I don't have the information to answer that question."

    Context:
    {context}

    Question: {query}
    Answer:"""

    import openai
    openai.api_key = LLM_API_KEY
    try:
        response = openai.Completion.create(
            model="gpt-3.5-turbo-instruct", # Choose an appropriate 
            prompt=prompt,
            max_tokens=200,
            temperature=0.2,
            n=1,
            stop=None
        )
        return response.choices[0].text.strip()
    except Exception as e:
        print(f"Error calling LLM: {e}")
        return "An error occurred while generating the response."

@app.route('/chat', methods=['POST'])
def chat():
    user_query = request.json.get('query')
    if not user_query:
        return jsonify({"error": "Missing query"}), 400

    # --- Personalization Layer (Conceptual) ---
    user_profile = get_user_profile(request.headers.get('Authorization')) # Example: Fetch user data
    personalized_context = get_personalized_context(user_profile) # Example: Fetch relevant account info

    # Augment query with personalized context (optional)
    augmented_query = f"{user_query} Regarding my {personalized_context}." if personalized_context else user_query

    relevant_documents = retrieve_relevant_documents(augmented_query)
    response = generate_response(user_query, relevant_documents)

    return jsonify({"response": response})

def get_user_profile(auth_token):
    # In a real application, you would authenticate the token and fetch user data
    # from your bank's user .
    # For this example, let's return a mock profile.
    if auth_token == "Bearer valid_token":
        return {"account_type": "checking", "recent_transactions": [...] }
    return None

def get_personalized_context(user_profile):
    if user_profile and user_profile.get("account_type"):
        return f"my {user_profile['account_type']} account"
    return None

if __name__ == '__main__':
    app.run(debug=True)

III. LLM Integration (within the Backend):

The generate_response function in the backend code snippet demonstrates the integration with an LLM (using OpenAI’s API as an example). You would replace "gpt-3.5-turbo-instruct" with your chosen model and handle the API interactions accordingly.

IV. Redis Integration (within the Backend):

The backend code shows how Redis is used for:

  • Storing Embeddings: The store_embeddings_in_redis section in the Knowledge Base Preparation.
  • Retrieving Relevant Documents: The retrieve_relevant_documents function uses Redis’s vector search capabilities to find the most similar document embeddings to the user’s query embedding.

V. React.js Front-End Development:

JavaScript

import React, { useState } from 'react';

function ChatAgent() {
  const [messages, setMessages] = useState([]);
  const [inputText, setInputText] = useState('');
  const [isLoading, setIsLoading] = useState(false);

  const sendMessage = async () => {
    if (!inputText.trim()) return;

    const userMessage = { text: inputText, sender: 'user' };
    setMessages([...messages, userMessage]);
    setInputText('');
    setIsLoading(true);

    try {
      const response = await fetch('/chat', {
        method: 'POST',
        headers: {
          'Content-Type': 'application/json',
          'Authorization': 'Bearer valid_token' // Example: Pass user token if authenticated
        },
        body: JSON.stringify({ query: inputText }),
      });

      if (!response.ok) {
        throw new Error(`HTTP error! status: ${response.status}`);
      }

      const data = await response.json();
      const botMessage = { text: data.response, sender: 'bot' };
      setMessages([...messages, botMessage]);
    } catch (error) {
      console.error("Error sending message:", error);
      const errorMessage = { text: "Sorry, I encountered an error.", sender: 'bot' };
      setMessages([...messages, errorMessage]);
    } finally {
      setIsLoading(false);
    }
  };

  return (
    <div className="chat-container">
      <div className="message-list">
        {messages.map((msg, index) => (
          <div key={index} className={`message ${msg.sender}`}>
            {msg.text}
          </div>
        ))}
        {isLoading && <div className="message bot">Loading...</div>}
      </div>
      <div className="input-area">
        <input
          type="text"
          value={inputText}
          onChange={(e) => setInputText(e.target.value)}
          placeholder="Ask a question..."
          onKeyPress={(e) => e.key === 'Enter' && sendMessage()}
        />
        <button onClick={sendMessage} disabled={isLoading}>Send</button>
      </div>
    </div>
  );
}

export default ChatAgent;

VI. Personalization Layer:

The personalization aspect is touched upon in the backend (/chat route and the get_user_profile, get_personalized_context functions). In a real-world scenario, this layer would involve:

  • User Authentication: Securely identifying the user.
  • Data Fetching: Retrieving relevant user data from your bank’s systems based on their identity (e.g., account details, transaction history, past interactions).
  • Contextualization: Using the fetched data to:
    • Filter/Boost Knowledge Base Results: Prioritize FAQs or document sections relevant to the user’s situation.
    • Augment the Query: Add context to the user’s query before retrieval (as shown in the backend example).
    • Tailor the Prompt: Include personalized information in the prompt sent to the LLM.

VII. Evaluation and Improvement:

This is an ongoing process that involves:

  • Tracking Metrics: Monitor user engagement, satisfaction, and the accuracy of the ‘s responses.
  • User Feedback Collection: Implement mechanisms for users to provide feedback on the chatbot’s answers.
  • Analysis: Analyze the data and feedback to identify areas where the chatbot can be improved (e.g., gaps in the knowledge base, poor-performing prompts).
  • Iteration: Continuously update the knowledge base, refine the RAG pipeline, and adjust the LLM prompts based on the evaluation results.

Important Considerations:

  • Security: Implement robust security measures at every layer, especially when handling user data and API keys.
  • Error Handling: Implement comprehensive error handling to gracefully manage unexpected issues.
  • Scalability: Design your system to handle a growing number of users and data.
  • Cost Management: Be mindful of the costs associated with LLM API usage and Redis hosting.
  • User Experience: Focus on creating a smooth and intuitive chat interface.
  • Compliance: Ensure your chatbot complies with all relevant banking regulations and privacy policies.

This detailed breakdown with sample code provides a solid foundation for building your personalized bank FAQ chat agent. Remember to adapt and expand upon this based on your specific requirements and the complexity of your bank’s information. Good luck!

Related Posts

More posts