Building Chatbots with Infactory

Chatbots and virtual assistants powered by Infactory can answer specific questions about your data accurately and efficiently. This guide explains how to build chatbots that leverage Infactory’s data intelligence capabilities.

Why Infactory for Chatbots?

Traditional AI-powered chatbots face several challenges when answering data-specific questions:

Hallucination

LLMs can make up false information when they don’t know the answer

Consistency

Answers to the same question may vary between sessions

Performance

Each query requires full LLM processing, which can be slow

Complexity

Requires complex prompt engineering and fine-tuning

Infactory solves these challenges by:

Executing queries directly against your data, ensuring accurate answers
Providing consistent results for the same questions
Processing queries at database speed, not AI inference speed
Simplifying development with an easy-to-use API

Architecture Overview

User asks a question

The user enters a natural language question in your chatbot interface.

Question sent to your backend

Your application sends the question to your backend server.

Backend calls Infactory API

Your backend server forwards the question to Infactory’s unified API endpoint.

Infactory processes the question

Infactory selects the appropriate query, extracts parameters, and executes it against your data.

Return structured data response

Infactory returns a structured data response to your backend.

Format and display the answer

Your application formats the structured data into a natural language response and displays it to the user.

Implementation Options

Simple Question-Answer Bot

The simplest implementation directly maps user questions to Infactory responses:

import { useState } from 'react';

function ChatBot() {
  const [messages, setMessages] = useState([]);
  const [input, setInput] = useState('');
  const [isLoading, setIsLoading] = useState(false);
  
  async function handleSubmit(e) {
    e.preventDefault();
    
    if (!input.trim()) return;
    
    // Add user message
    const userMessage = { role: 'user', content: input };
    setMessages(prev => [...prev, userMessage]);
    setInput('');
    setIsLoading(true);
    
    try {
      // Send question to backend
      const response = await fetch('/api/chat', {
        method: 'POST',
        headers: { 'Content-Type': 'application/json' },
        body: JSON.stringify({ question: userMessage.content })
      });
      
      const data = await response.json();
      
      // Format the response
      let botResponse;
      if (data.error) {
        botResponse = { role: 'assistant', content: `I'm sorry, I couldn't find an answer. ${data.error}` };
      } else {
        // Create a natural language response from the structured data
        botResponse = { 
          role: 'assistant', 
          content: formatResponse(data),
          data: data.data // Store the raw data for display if needed
        };
      }
      
      setMessages(prev => [...prev, botResponse]);
    } catch (error) {
      setMessages(prev => [
        ...prev, 
        { role: 'assistant', content: "I'm sorry, I encountered an error. Please try again." }
      ]);
    } finally {
      setIsLoading(false);
    }
  }
  
  // Format structured data into a natural language response
  function formatResponse(data) {
    if (!data.data || data.data.length === 0) {
      return "I couldn't find any information about that.";
    }
    
    // Example formatting for average by category query
    if (data.query_used === 'average_by_category') {
      const results = data.data.map(item => 
        `The average ${data.parameters.metric} for ${item[data.parameters.category]} is ${item.average.toFixed(2)}`
      ).join('. ');
      
      return results;
    }
    
    // Default response if we don't have specific formatting for this query type
    return `Here's what I found: ${JSON.stringify(data.data)}`;
  }
  
  return (
    <div className="chatbot">
      <div className="messages">
        {messages.map((msg, index) => (
          <div key={index} className={`message ${msg.role}`}>
            <div className="content">{msg.content}</div>
            {msg.data && (
              <button 
                onClick={() => console.log(msg.data)} 
                className="view-data-btn"
              >
                View Data
              </button>
            )}
          </div>
        ))}
        {isLoading && (
          <div className="message assistant loading">
            <div className="typing-indicator">
              <span></span><span></span><span></span>
            </div>
          </div>
        )}
      </div>
      
      <form onSubmit={handleSubmit} className="input-form">
        <input
          type="text"
          value={input}
          onChange={(e) => setInput(e.target.value)}
          placeholder="Ask a question about your data..."
          disabled={isLoading}
        />
        <button type="submit" disabled={isLoading}>
          Send
        </button>
      </form>
    </div>
  );
}

import { useState } from 'react';

function ChatBot() {
  const [messages, setMessages] = useState([]);
  const [input, setInput] = useState('');
  const [isLoading, setIsLoading] = useState(false);
  
  async function handleSubmit(e) {
    e.preventDefault();
    
    if (!input.trim()) return;
    
    // Add user message
    const userMessage = { role: 'user', content: input };
    setMessages(prev => [...prev, userMessage]);
    setInput('');
    setIsLoading(true);
    
    try {
      // Send question to backend
      const response = await fetch('/api/chat', {
        method: 'POST',
        headers: { 'Content-Type': 'application/json' },
        body: JSON.stringify({ question: userMessage.content })
      });
      
      const data = await response.json();
      
      // Format the response
      let botResponse;
      if (data.error) {
        botResponse = { role: 'assistant', content: `I'm sorry, I couldn't find an answer. ${data.error}` };
      } else {
        // Create a natural language response from the structured data
        botResponse = { 
          role: 'assistant', 
          content: formatResponse(data),
          data: data.data // Store the raw data for display if needed
        };
      }
      
      setMessages(prev => [...prev, botResponse]);
    } catch (error) {
      setMessages(prev => [
        ...prev, 
        { role: 'assistant', content: "I'm sorry, I encountered an error. Please try again." }
      ]);
    } finally {
      setIsLoading(false);
    }
  }
  
  // Format structured data into a natural language response
  function formatResponse(data) {
    if (!data.data || data.data.length === 0) {
      return "I couldn't find any information about that.";
    }
    
    // Example formatting for average by category query
    if (data.query_used === 'average_by_category') {
      const results = data.data.map(item => 
        `The average ${data.parameters.metric} for ${item[data.parameters.category]} is ${item.average.toFixed(2)}`
      ).join('. ');
      
      return results;
    }
    
    // Default response if we don't have specific formatting for this query type
    return `Here's what I found: ${JSON.stringify(data.data)}`;
  }
  
  return (
    <div className="chatbot">
      <div className="messages">
        {messages.map((msg, index) => (
          <div key={index} className={`message ${msg.role}`}>
            <div className="content">{msg.content}</div>
            {msg.data && (
              <button 
                onClick={() => console.log(msg.data)} 
                className="view-data-btn"
              >
                View Data
              </button>
            )}
          </div>
        ))}
        {isLoading && (
          <div className="message assistant loading">
            <div className="typing-indicator">
              <span></span><span></span><span></span>
            </div>
          </div>
        )}
      </div>
      
      <form onSubmit={handleSubmit} className="input-form">
        <input
          type="text"
          value={input}
          onChange={(e) => setInput(e.target.value)}
          placeholder="Ask a question about your data..."
          disabled={isLoading}
        />
        <button type="submit" disabled={isLoading}>
          Send
        </button>
      </form>
    </div>
  );
}

const express = require('express');
const axios = require('axios');
const router = express.Router();

router.post('/api/chat', async (req, res) => {
  try {
    const { question } = req.body;
    
    // Call Infactory API
    const response = await axios.post(
      'https://api.infactory.ai/v1/query',
      {
        query: question,
        project_id: process.env.INFACTORY_PROJECT_ID,
      },
      {
        headers: {
          'Content-Type': 'application/json',
          'Authorization': `Bearer ${process.env.INFACTORY_API_KEY}`
        }
      }
    );
    
    res.json(response.data);
  } catch (error) {
    console.error('Error querying Infactory:', error.response?.data || error.message);
    
    let errorMessage = 'Failed to query the database';
    
    // Handle specific error types
    if (error.response?.data?.error === 'no_matching_query') {
      errorMessage = "I don't know how to answer that question yet.";
    }
    
    res.status(500).json({ error: errorMessage });
  }
});

module.exports = router;

Hybrid LLM + Infactory Chatbot

For a more conversational experience, you can combine Infactory with an LLM:

In this approach:

The LLM handles conversation flow and identifies when data questions are asked
Infactory answers specific data questions
The LLM formats the responses naturally

const express = require('express');
const axios = require('axios');
const { OpenAI } = require('openai');
const router = express.Router();

const openai = new OpenAI({
  apiKey: process.env.OPENAI_API_KEY
});

router.post('/api/chat', async (req, res) => {
  try {
    const { question, history } = req.body;
    
    // Step 1: Have the LLM determine if this is a data question
    const analysisPrompt = `
    You are an assistant that helps determine if a question requires database querying.
    User question: "${question}"
    
    Is this a question that requires querying a database to answer correctly? 
    Respond with YES if it's asking for specific data, statistics, or information that would be stored in a database.
    Respond with NO if it's a general question, chitchat, or something that doesn't require looking up specific data.
    Just respond with YES or NO.`;
    
    const analysisResponse = await openai.chat.completions.create({
      model: "gpt-3.5-turbo",
      messages: [{ role: "user", content: analysisPrompt }],
      temperature: 0,
    });
    
    const isDataQuestion = analysisResponse.choices[0].message.content.trim().startsWith('YES');
    
    let responseData;
    
    if (isDataQuestion) {
      // Step 2: If it's a data question, query Infactory
      const infactoryResponse = await axios.post(
        'https://api.infactory.ai/v1/query',
        {
          query: question,
          project_id: process.env.INFACTORY_PROJECT_ID,
        },
        {
          headers: {
            'Content-Type': 'application/json',
            'Authorization': `Bearer ${process.env.INFACTORY_API_KEY}`
          }
        }
      );
      
      // Step 3: Have the LLM format the response
      const formattingPrompt = `
      You are a helpful assistant. Format the following data into a natural, conversational response.
      User question: "${question}"
      Data: ${JSON.stringify(infactoryResponse.data)}
      
      Provide a conversational, easy to understand answer based strictly on this data.`;
      
      const formattingResponse = await openai.chat.completions.create({
        model: "gpt-3.5-turbo",
        messages: [{ role: "user", content: formattingPrompt }],
        temperature: 0.7,
      });
      
      responseData = {
        content: formattingResponse.choices[0].message.content,
        data: infactoryResponse.data.data,
        source: 'Infactory'
      };
    } else {
      // Step 4: If it's not a data question, let the LLM handle it
      const chatMessages = [
        { role: "system", content: "You are a helpful assistant that works with a database system. For data-specific questions, those are handled by another system. Focus on being helpful for general questions, clarifications, and conversation." },
        ...history.map(msg => ({ role: msg.role, content: msg.content })),
        { role: "user", content: question }
      ];
      
      const chatResponse = await openai.chat.completions.create({
        model: "gpt-3.5-turbo",
        messages: chatMessages,
        temperature: 0.7,
      });
      
      responseData = {
        content: chatResponse.choices[0].message.content,
        source: 'llm'
      };
    }
    
    res.json(responseData);
  } catch (error) {
    console.error('Chatbot error:', error);
    res.status(500).json({ error: 'Failed to process your request' });
  }
});

module.exports = router;

const express = require('express');
const axios = require('axios');
const { OpenAI } = require('openai');
const router = express.Router();

const openai = new OpenAI({
  apiKey: process.env.OPENAI_API_KEY
});

router.post('/api/chat', async (req, res) => {
  try {
    const { question, history } = req.body;
    
    // Step 1: Have the LLM determine if this is a data question
    const analysisPrompt = `
    You are an assistant that helps determine if a question requires database querying.
    User question: "${question}"
    
    Is this a question that requires querying a database to answer correctly? 
    Respond with YES if it's asking for specific data, statistics, or information that would be stored in a database.
    Respond with NO if it's a general question, chitchat, or something that doesn't require looking up specific data.
    Just respond with YES or NO.`;
    
    const analysisResponse = await openai.chat.completions.create({
      model: "gpt-3.5-turbo",
      messages: [{ role: "user", content: analysisPrompt }],
      temperature: 0,
    });
    
    const isDataQuestion = analysisResponse.choices[0].message.content.trim().startsWith('YES');
    
    let responseData;
    
    if (isDataQuestion) {
      // Step 2: If it's a data question, query Infactory
      const infactoryResponse = await axios.post(
        'https://api.infactory.ai/v1/query',
        {
          query: question,
          project_id: process.env.INFACTORY_PROJECT_ID,
        },
        {
          headers: {
            'Content-Type': 'application/json',
            'Authorization': `Bearer ${process.env.INFACTORY_API_KEY}`
          }
        }
      );
      
      // Step 3: Have the LLM format the response
      const formattingPrompt = `
      You are a helpful assistant. Format the following data into a natural, conversational response.
      User question: "${question}"
      Data: ${JSON.stringify(infactoryResponse.data)}
      
      Provide a conversational, easy to understand answer based strictly on this data.`;
      
      const formattingResponse = await openai.chat.completions.create({
        model: "gpt-3.5-turbo",
        messages: [{ role: "user", content: formattingPrompt }],
        temperature: 0.7,
      });
      
      responseData = {
        content: formattingResponse.choices[0].message.content,
        data: infactoryResponse.data.data,
        source: 'Infactory'
      };
    } else {
      // Step 4: If it's not a data question, let the LLM handle it
      const chatMessages = [
        { role: "system", content: "You are a helpful assistant that works with a database system. For data-specific questions, those are handled by another system. Focus on being helpful for general questions, clarifications, and conversation." },
        ...history.map(msg => ({ role: msg.role, content: msg.content })),
        { role: "user", content: question }
      ];
      
      const chatResponse = await openai.chat.completions.create({
        model: "gpt-3.5-turbo",
        messages: chatMessages,
        temperature: 0.7,
      });
      
      responseData = {
        content: chatResponse.choices[0].message.content,
        source: 'llm'
      };
    }
    
    res.json(responseData);
  } catch (error) {
    console.error('Chatbot error:', error);
    res.status(500).json({ error: 'Failed to process your request' });
  }
});

module.exports = router;

Enhancing Your Chatbot

Support for Follow-up Questions

To handle follow-up questions, maintain conversation context and reference previous responses:

// Track the last query and its parameters
let lastQuery = null;

// Example follow-up handler
function handleFollowup(question, history) {
  const lastMessage = history[history.length - 2]; // Get the last assistant message
  
  if (lastMessage && lastMessage.data && question.toLowerCase().includes('what about')) {
    // This looks like a follow-up question
    if (lastMessage.queryInfo.query_used === 'average_by_category') {
      // Extract what they're asking about
      const match = question.match(/what about (the |)([a-z0-9 ]+)/i);
      if (match) {
        const newCategory = match[2].trim();
        
        // Update the parameters from the last query
        return {
          isFollowup: true,
          queryName: lastMessage.queryInfo.query_used,
          parameters: {
            ...lastMessage.queryInfo.parameters,
            category: newCategory
          }
        };
      }
    }
  }
  
  return { isFollowup: false };
}

Adding Visualizations

Enhance your chatbot with visual representations of data:

import { useState, useRef } from 'react';
import { Bar, Line, Pie } from 'react-chartjs-2';
import { Chart as ChartJS, CategoryScale, LinearScale, BarElement, PointElement, LineElement, ArcElement, Title, Tooltip, Legend } from 'chart.js';

// Register Chart.js components
ChartJS.register(CategoryScale, LinearScale, BarElement, PointElement, LineElement, ArcElement, Title, Tooltip, Legend);

function ChatBotMessage({ message }) {
  const [showChart, setShowChart] = useState(false);
  
  // Only process data visualizations for assistant messages with data
  if (message.role !== 'assistant' || !message.data) {
    return (
      <div className={`message ${message.role}`}>
        <div className="content">{message.content}</div>
      </div>
    );
  }
  
  // Determine what type of chart to show based on the data
  const getChartType = () => {
    if (!message.data || !Array.isArray(message.data)) return null;
    
    // Example: For data that has categories and a single metric, use a bar chart
    if (message.data.length > 1 && message.data.length <= 10) {
      const firstItem = message.data[0];
      const keys = Object.keys(firstItem);
      
      if (keys.length === 2 && typeof firstItem[keys[1]] === 'number') {
        // Likely a category and a value - good for a bar chart
        return 'bar';
      } else if (keys.includes('date') || keys.includes('time') || keys.some(k => k.includes('date') || k.includes('time'))) {
        // Time series data - good for a line chart
        return 'line';
      } else if (message.data.length <= 6) {
        // Small number of categories - could be good for a pie chart
        return 'pie';
      }
    }
    
    return 'bar'; // Default to bar chart
  };
  
  const chartType = getChartType();
  
  // Prepare chart data based on the data structure
  const prepareChartData = () => {
    if (!message.data || !Array.isArray(message.data) || message.data.length === 0) return null;
    
    const firstItem = message.data[0];
    const keys = Object.keys(firstItem);
    
    // For a simple two-column data structure (category and value)
    if (keys.length === 2 && typeof firstItem[keys[1]] === 'number') {
      const categoryKey = keys[0];
      const valueKey = keys[1];
      
      return {
        labels: message.data.map(item => item[categoryKey]),
        datasets: [
          {
            label: valueKey.charAt(0).toUpperCase() + valueKey.slice(1),
            data: message.data.map(item => item[valueKey]),
            backgroundColor: 'rgba(53, 162, 235, 0.5)',
            borderColor: 'rgba(53, 162, 235, 1)',
            borderWidth: 1
          }
        ]
      };
    }
    
    // Add more data preparation logic for other data structures
    
    return null;
  };
  
  const chartData = prepareChartData();
  
  // Chart options
  const chartOptions = {
    responsive: true,
    plugins: {
      legend: {
        position: 'top',
      },
      title: {
        display: true,
        text: 'Data Visualization',
      },
    },
  };
  
  // Render the appropriate chart component based on chartType
  const renderChart = () => {
    if (!chartData) return null;
    
    switch (chartType) {
      case 'bar':
        return <Bar data={chartData} options={chartOptions} />;
      case 'line':
        return <Line data={chartData} options={chartOptions} />;
      case 'pie':
        return <Pie data={chartData} options={chartOptions} />;
      default:
        return null;
    }
  };
  
  return (
    <div className={`message ${message.role}`}>
      <div className="content">{message.content}</div>
      
      {chartType && chartData && (
        <div className="message-actions">
          <button onClick={() => setShowChart(!showChart)}>
            {showChart ? 'Hide Visualization' : 'Show Visualization'}
          </button>
          
          {showChart && (
            <div className="chart-container">
              {renderChart()}
            </div>
          )}
        </div>
      )}
      
      <div className="message-footer">
        <small>Source: Infactory data query</small>
      </div>
    </div>
  );
}

import { useState, useRef } from 'react';
import { Bar, Line, Pie } from 'react-chartjs-2';
import { Chart as ChartJS, CategoryScale, LinearScale, BarElement, PointElement, LineElement, ArcElement, Title, Tooltip, Legend } from 'chart.js';

// Register Chart.js components
ChartJS.register(CategoryScale, LinearScale, BarElement, PointElement, LineElement, ArcElement, Title, Tooltip, Legend);

function ChatBotMessage({ message }) {
  const [showChart, setShowChart] = useState(false);
  
  // Only process data visualizations for assistant messages with data
  if (message.role !== 'assistant' || !message.data) {
    return (
      <div className={`message ${message.role}`}>
        <div className="content">{message.content}</div>
      </div>
    );
  }
  
  // Determine what type of chart to show based on the data
  const getChartType = () => {
    if (!message.data || !Array.isArray(message.data)) return null;
    
    // Example: For data that has categories and a single metric, use a bar chart
    if (message.data.length > 1 && message.data.length <= 10) {
      const firstItem = message.data[0];
      const keys = Object.keys(firstItem);
      
      if (keys.length === 2 && typeof firstItem[keys[1]] === 'number') {
        // Likely a category and a value - good for a bar chart
        return 'bar';
      } else if (keys.includes('date') || keys.includes('time') || keys.some(k => k.includes('date') || k.includes('time'))) {
        // Time series data - good for a line chart
        return 'line';
      } else if (message.data.length <= 6) {
        // Small number of categories - could be good for a pie chart
        return 'pie';
      }
    }
    
    return 'bar'; // Default to bar chart
  };
  
  const chartType = getChartType();
  
  // Prepare chart data based on the data structure
  const prepareChartData = () => {
    if (!message.data || !Array.isArray(message.data) || message.data.length === 0) return null;
    
    const firstItem = message.data[0];
    const keys = Object.keys(firstItem);
    
    // For a simple two-column data structure (category and value)
    if (keys.length === 2 && typeof firstItem[keys[1]] === 'number') {
      const categoryKey = keys[0];
      const valueKey = keys[1];
      
      return {
        labels: message.data.map(item => item[categoryKey]),
        datasets: [
          {
            label: valueKey.charAt(0).toUpperCase() + valueKey.slice(1),
            data: message.data.map(item => item[valueKey]),
            backgroundColor: 'rgba(53, 162, 235, 0.5)',
            borderColor: 'rgba(53, 162, 235, 1)',
            borderWidth: 1
          }
        ]
      };
    }
    
    // Add more data preparation logic for other data structures
    
    return null;
  };
  
  const chartData = prepareChartData();
  
  // Chart options
  const chartOptions = {
    responsive: true,
    plugins: {
      legend: {
        position: 'top',
      },
      title: {
        display: true,
        text: 'Data Visualization',
      },
    },
  };
  
  // Render the appropriate chart component based on chartType
  const renderChart = () => {
    if (!chartData) return null;
    
    switch (chartType) {
      case 'bar':
        return <Bar data={chartData} options={chartOptions} />;
      case 'line':
        return <Line data={chartData} options={chartOptions} />;
      case 'pie':
        return <Pie data={chartData} options={chartOptions} />;
      default:
        return null;
    }
  };
  
  return (
    <div className={`message ${message.role}`}>
      <div className="content">{message.content}</div>
      
      {chartType && chartData && (
        <div className="message-actions">
          <button onClick={() => setShowChart(!showChart)}>
            {showChart ? 'Hide Visualization' : 'Show Visualization'}
          </button>
          
          {showChart && (
            <div className="chart-container">
              {renderChart()}
            </div>
          )}
        </div>
      )}
      
      <div className="message-footer">
        <small>Source: Infactory data query</small>
      </div>
    </div>
  );
}

Multi-turn Context Awareness

To handle complex conversations with multiple turns:

// Conversation state manager
class ConversationState {
  constructor() {
    this.topics = new Map(); // Track topics mentioned in the conversation
    this.lastQuery = null;   // Last executed query
    this.lastResponse = null; // Last response data
    this.context = {};       // Additional context like filters applied
  }
  
  updateFromMessage(message, isUser) {
    if (isUser) {
      // Process user message to extract topics, entities, etc.
      this.extractTopics(message);
    } else {
      // Update based on system response
      if (message.queryInfo) {
        this.lastQuery = message.queryInfo.query_used;
        this.lastResponse = message.data;
        
        // Update context based on the parameters used
        if (message.queryInfo.parameters) {
          Object.entries(message.queryInfo.parameters).forEach(([key, value]) => {
            this.context[key] = value;
          });
        }
      }
    }
  }
  
  extractTopics(message) {
    // Simple implementation - in production you might use NLP services
    const lowerMessage = message.toLowerCase();
    
    // Check for key topics relevant to your data domain
    const potentialTopics = ['sales', 'revenue', 'customers', 'products', 'regions'];
    
    potentialTopics.forEach(topic => {
      if (lowerMessage.includes(topic)) {
        // Increase the relevance score for this topic
        const currentScore = this.topics.get(topic) || 0;
        this.topics.set(topic, currentScore + 1);
      }
    });
  }
  
  getRelevantContext(message) {
    // Determine what context is relevant to the current message
    const context = { ...this.context };
    
    // Decay or remove irrelevant context based on the new message
    // This is highly domain-specific
    
    return context;
  }
}

Deployment Considerations

Scaling Your Chatbot

As usage grows, consider these scaling strategies:

Load Balancing

Distribute API calls across multiple backend instances

Caching

Cache common queries to reduce API calls and improve response times

Queue Processing

Use a message queue for handling high volumes of requests

Serverless Functions

Deploy as serverless functions that scale automatically with demand

Monitoring and Analytics

Track your chatbot’s performance:

Usage Metrics: Number of questions, unique users, peak usage times
Performance Metrics: Response times, API call success rates
Content Metrics: Most common questions, unanswered questions
User Satisfaction Metrics: Feedback ratings, abandonment rates

Best Practices

Clear Scope

Define what types of questions your chatbot should answer

Fallback Strategies

Have graceful responses for questions outside your query coverage

Progressive Enhancement

Start simple and add complexity as you learn user patterns

Continuous Improvement

Regularly analyze unanswerable questions to identify new queries to create

User Feedback Loop

Allow users to flag incorrect or unhelpful answers

Test with Real Users

Validate with actual users rather than just hypothetical questions

Common Challenges and Solutions

Handling Ambiguous Questions

Challenge: Users ask questions that could match multiple queries or have unclear parameters.

Solution: Implement a clarification flow where the chatbot asks follow-up questions when a question is ambiguous. For example:

if (matchingQueries.length > 1 && matchingQueries[0].confidence < 0.8) {
  return {
    needsClarification: true,
    clarificationQuestion: `I'm not sure exactly what you're asking about. Are you interested in ${matchingQueries.map(q => q.topic).join(' or ')}?`,
    matchingQueries: matchingQueries
  };
}

Maintaining Conversation Context

Challenge: Keeping track of context across multiple turns of conversation.

Solution: Implement a context management system that tracks entities, topics, and previous queries. Use this context to enhance subsequent queries.

// Update context based on the current query
function updateContext(context, query, parameters) {
  return {
    ...context,
    lastQuery: query,
    entities: {
      ...context.entities,
      ...extractEntities(parameters)
    },
    timestamp: Date.now()
  };
}

// Apply context to a new question
function applyContext(question, context) {
  // If the question includes pronouns like "it", "them", "those", 
  // try to resolve using context
  if (/\b(it|them|those|that|these)\b/i.test(question)) {
    if (context.entities.category) {
      // Replace pronouns with the actual entity from context
      return question.replace(
        /\b(it|them|those|that|these)\b/i, 
        context.entities.category
      );
    }
  }
  
  return question;
}

Unrealistic User Expectations

Challenge: Users expect the chatbot to answer any question, regardless of whether the data supports it.

Solution: Set clear expectations and provide informative responses about the chatbot’s capabilities:

function handleUnanswerable(question) {
  // Check if the question is about a topic we don't have data for
  const unsupportedTopics = ['employee salaries', 'future predictions', 'competitor data'];
  
  for (const topic of unsupportedTopics) {
    if (question.toLowerCase().includes(topic)) {
      return `I don't have information about ${topic}. I can help you with questions about our sales, customers, products, and regional performance.`;
    }
  }
  
  // General fallback
  return "I don't have enough information to answer that question. I can help you with questions about our business data such as sales figures, customer metrics, and product performance.";
}

Example Use Cases

Customer Service Bot

Answer questions about order status, shipping times, and product availability

Internal Analytics Assistant

Help employees explore business metrics without needing SQL knowledge

Sales Dashboard Companion

Provide conversational access to sales performance data

Product Recommendation Bot

Recommend products based on customer data and preferences

Next Steps

After building your chatbot, consider:

Implementing user feedback mechanisms to improve your queries
Adding authentication and authorization for internal chatbots
Exploring other integration patterns like dashboards or search interfaces
Setting up performance monitoring to ensure responsiveness

On this page

Building Chatbots with Infactory
Why Infactory for Chatbots?
Architecture Overview
Implementation Options
Simple Question-Answer Bot
Hybrid LLM + Infactory Chatbot
Enhancing Your Chatbot
Support for Follow-up Questions
Adding Visualizations
Multi-turn Context Awareness
Deployment Considerations
Scaling Your Chatbot
Monitoring and Analytics
Best Practices
Common Challenges and Solutions
Example Use Cases
Next Steps

Building Chatbots with Infactory

Why Infactory for Chatbots?

Traditional AI-powered chatbots face several challenges when answering data-specific questions:

Hallucination

LLMs can make up false information when they don’t know the answer

Consistency

Answers to the same question may vary between sessions

Performance

Each query requires full LLM processing, which can be slow

Complexity

Requires complex prompt engineering and fine-tuning

Infactory solves these challenges by:

Executing queries directly against your data, ensuring accurate answers
Providing consistent results for the same questions
Processing queries at database speed, not AI inference speed
Simplifying development with an easy-to-use API

Architecture Overview

User asks a question

The user enters a natural language question in your chatbot interface.

Question sent to your backend

Your application sends the question to your backend server.

Backend calls Infactory API

Your backend server forwards the question to Infactory’s unified API endpoint.

Infactory processes the question

Infactory selects the appropriate query, extracts parameters, and executes it against your data.

Return structured data response

Infactory returns a structured data response to your backend.

Format and display the answer

Your application formats the structured data into a natural language response and displays it to the user.

Implementation Options

Simple Question-Answer Bot

The simplest implementation directly maps user questions to Infactory responses:

import { useState } from 'react';

function ChatBot() {
  const [messages, setMessages] = useState([]);
  const [input, setInput] = useState('');
  const [isLoading, setIsLoading] = useState(false);
  
  async function handleSubmit(e) {
    e.preventDefault();
    
    if (!input.trim()) return;
    
    // Add user message
    const userMessage = { role: 'user', content: input };
    setMessages(prev => [...prev, userMessage]);
    setInput('');
    setIsLoading(true);
    
    try {
      // Send question to backend
      const response = await fetch('/api/chat', {
        method: 'POST',
        headers: { 'Content-Type': 'application/json' },
        body: JSON.stringify({ question: userMessage.content })
      });
      
      const data = await response.json();
      
      // Format the response
      let botResponse;
      if (data.error) {
        botResponse = { role: 'assistant', content: `I'm sorry, I couldn't find an answer. ${data.error}` };
      } else {
        // Create a natural language response from the structured data
        botResponse = { 
          role: 'assistant', 
          content: formatResponse(data),
          data: data.data // Store the raw data for display if needed
        };
      }
      
      setMessages(prev => [...prev, botResponse]);
    } catch (error) {
      setMessages(prev => [
        ...prev, 
        { role: 'assistant', content: "I'm sorry, I encountered an error. Please try again." }
      ]);
    } finally {
      setIsLoading(false);
    }
  }
  
  // Format structured data into a natural language response
  function formatResponse(data) {
    if (!data.data || data.data.length === 0) {
      return "I couldn't find any information about that.";
    }
    
    // Example formatting for average by category query
    if (data.query_used === 'average_by_category') {
      const results = data.data.map(item => 
        `The average ${data.parameters.metric} for ${item[data.parameters.category]} is ${item.average.toFixed(2)}`
      ).join('. ');
      
      return results;
    }
    
    // Default response if we don't have specific formatting for this query type
    return `Here's what I found: ${JSON.stringify(data.data)}`;
  }
  
  return (
    <div className="chatbot">
      <div className="messages">
        {messages.map((msg, index) => (
          <div key={index} className={`message ${msg.role}`}>
            <div className="content">{msg.content}</div>
            {msg.data && (
              <button 
                onClick={() => console.log(msg.data)} 
                className="view-data-btn"
              >
                View Data
              </button>
            )}
          </div>
        ))}
        {isLoading && (
          <div className="message assistant loading">
            <div className="typing-indicator">
              <span></span><span></span><span></span>
            </div>
          </div>
        )}
      </div>
      
      <form onSubmit={handleSubmit} className="input-form">
        <input
          type="text"
          value={input}
          onChange={(e) => setInput(e.target.value)}
          placeholder="Ask a question about your data..."
          disabled={isLoading}
        />
        <button type="submit" disabled={isLoading}>
          Send
        </button>
      </form>
    </div>
  );
}

import { useState } from 'react';

function ChatBot() {
  const [messages, setMessages] = useState([]);
  const [input, setInput] = useState('');
  const [isLoading, setIsLoading] = useState(false);
  
  async function handleSubmit(e) {
    e.preventDefault();
    
    if (!input.trim()) return;
    
    // Add user message
    const userMessage = { role: 'user', content: input };
    setMessages(prev => [...prev, userMessage]);
    setInput('');
    setIsLoading(true);
    
    try {
      // Send question to backend
      const response = await fetch('/api/chat', {
        method: 'POST',
        headers: { 'Content-Type': 'application/json' },
        body: JSON.stringify({ question: userMessage.content })
      });
      
      const data = await response.json();
      
      // Format the response
      let botResponse;
      if (data.error) {
        botResponse = { role: 'assistant', content: `I'm sorry, I couldn't find an answer. ${data.error}` };
      } else {
        // Create a natural language response from the structured data
        botResponse = { 
          role: 'assistant', 
          content: formatResponse(data),
          data: data.data // Store the raw data for display if needed
        };
      }
      
      setMessages(prev => [...prev, botResponse]);
    } catch (error) {
      setMessages(prev => [
        ...prev, 
        { role: 'assistant', content: "I'm sorry, I encountered an error. Please try again." }
      ]);
    } finally {
      setIsLoading(false);
    }
  }
  
  // Format structured data into a natural language response
  function formatResponse(data) {
    if (!data.data || data.data.length === 0) {
      return "I couldn't find any information about that.";
    }
    
    // Example formatting for average by category query
    if (data.query_used === 'average_by_category') {
      const results = data.data.map(item => 
        `The average ${data.parameters.metric} for ${item[data.parameters.category]} is ${item.average.toFixed(2)}`
      ).join('. ');
      
      return results;
    }
    
    // Default response if we don't have specific formatting for this query type
    return `Here's what I found: ${JSON.stringify(data.data)}`;
  }
  
  return (
    <div className="chatbot">
      <div className="messages">
        {messages.map((msg, index) => (
          <div key={index} className={`message ${msg.role}`}>
            <div className="content">{msg.content}</div>
            {msg.data && (
              <button 
                onClick={() => console.log(msg.data)} 
                className="view-data-btn"
              >
                View Data
              </button>
            )}
          </div>
        ))}
        {isLoading && (
          <div className="message assistant loading">
            <div className="typing-indicator">
              <span></span><span></span><span></span>
            </div>
          </div>
        )}
      </div>
      
      <form onSubmit={handleSubmit} className="input-form">
        <input
          type="text"
          value={input}
          onChange={(e) => setInput(e.target.value)}
          placeholder="Ask a question about your data..."
          disabled={isLoading}
        />
        <button type="submit" disabled={isLoading}>
          Send
        </button>
      </form>
    </div>
  );
}

const express = require('express');
const axios = require('axios');
const router = express.Router();

router.post('/api/chat', async (req, res) => {
  try {
    const { question } = req.body;
    
    // Call Infactory API
    const response = await axios.post(
      'https://api.infactory.ai/v1/query',
      {
        query: question,
        project_id: process.env.INFACTORY_PROJECT_ID,
      },
      {
        headers: {
          'Content-Type': 'application/json',
          'Authorization': `Bearer ${process.env.INFACTORY_API_KEY}`
        }
      }
    );
    
    res.json(response.data);
  } catch (error) {
    console.error('Error querying Infactory:', error.response?.data || error.message);
    
    let errorMessage = 'Failed to query the database';
    
    // Handle specific error types
    if (error.response?.data?.error === 'no_matching_query') {
      errorMessage = "I don't know how to answer that question yet.";
    }
    
    res.status(500).json({ error: errorMessage });
  }
});

module.exports = router;

Hybrid LLM + Infactory Chatbot

For a more conversational experience, you can combine Infactory with an LLM:

In this approach:

The LLM handles conversation flow and identifies when data questions are asked
Infactory answers specific data questions
The LLM formats the responses naturally

const express = require('express');
const axios = require('axios');
const { OpenAI } = require('openai');
const router = express.Router();

const openai = new OpenAI({
  apiKey: process.env.OPENAI_API_KEY
});

router.post('/api/chat', async (req, res) => {
  try {
    const { question, history } = req.body;
    
    // Step 1: Have the LLM determine if this is a data question
    const analysisPrompt = `
    You are an assistant that helps determine if a question requires database querying.
    User question: "${question}"
    
    Is this a question that requires querying a database to answer correctly? 
    Respond with YES if it's asking for specific data, statistics, or information that would be stored in a database.
    Respond with NO if it's a general question, chitchat, or something that doesn't require looking up specific data.
    Just respond with YES or NO.`;
    
    const analysisResponse = await openai.chat.completions.create({
      model: "gpt-3.5-turbo",
      messages: [{ role: "user", content: analysisPrompt }],
      temperature: 0,
    });
    
    const isDataQuestion = analysisResponse.choices[0].message.content.trim().startsWith('YES');
    
    let responseData;
    
    if (isDataQuestion) {
      // Step 2: If it's a data question, query Infactory
      const infactoryResponse = await axios.post(
        'https://api.infactory.ai/v1/query',
        {
          query: question,
          project_id: process.env.INFACTORY_PROJECT_ID,
        },
        {
          headers: {
            'Content-Type': 'application/json',
            'Authorization': `Bearer ${process.env.INFACTORY_API_KEY}`
          }
        }
      );
      
      // Step 3: Have the LLM format the response
      const formattingPrompt = `
      You are a helpful assistant. Format the following data into a natural, conversational response.
      User question: "${question}"
      Data: ${JSON.stringify(infactoryResponse.data)}
      
      Provide a conversational, easy to understand answer based strictly on this data.`;
      
      const formattingResponse = await openai.chat.completions.create({
        model: "gpt-3.5-turbo",
        messages: [{ role: "user", content: formattingPrompt }],
        temperature: 0.7,
      });
      
      responseData = {
        content: formattingResponse.choices[0].message.content,
        data: infactoryResponse.data.data,
        source: 'Infactory'
      };
    } else {
      // Step 4: If it's not a data question, let the LLM handle it
      const chatMessages = [
        { role: "system", content: "You are a helpful assistant that works with a database system. For data-specific questions, those are handled by another system. Focus on being helpful for general questions, clarifications, and conversation." },
        ...history.map(msg => ({ role: msg.role, content: msg.content })),
        { role: "user", content: question }
      ];
      
      const chatResponse = await openai.chat.completions.create({
        model: "gpt-3.5-turbo",
        messages: chatMessages,
        temperature: 0.7,
      });
      
      responseData = {
        content: chatResponse.choices[0].message.content,
        source: 'llm'
      };
    }
    
    res.json(responseData);
  } catch (error) {
    console.error('Chatbot error:', error);
    res.status(500).json({ error: 'Failed to process your request' });
  }
});

module.exports = router;

const express = require('express');
const axios = require('axios');
const { OpenAI } = require('openai');
const router = express.Router();

const openai = new OpenAI({
  apiKey: process.env.OPENAI_API_KEY
});

router.post('/api/chat', async (req, res) => {
  try {
    const { question, history } = req.body;
    
    // Step 1: Have the LLM determine if this is a data question
    const analysisPrompt = `
    You are an assistant that helps determine if a question requires database querying.
    User question: "${question}"
    
    Is this a question that requires querying a database to answer correctly? 
    Respond with YES if it's asking for specific data, statistics, or information that would be stored in a database.
    Respond with NO if it's a general question, chitchat, or something that doesn't require looking up specific data.
    Just respond with YES or NO.`;
    
    const analysisResponse = await openai.chat.completions.create({
      model: "gpt-3.5-turbo",
      messages: [{ role: "user", content: analysisPrompt }],
      temperature: 0,
    });
    
    const isDataQuestion = analysisResponse.choices[0].message.content.trim().startsWith('YES');
    
    let responseData;
    
    if (isDataQuestion) {
      // Step 2: If it's a data question, query Infactory
      const infactoryResponse = await axios.post(
        'https://api.infactory.ai/v1/query',
        {
          query: question,
          project_id: process.env.INFACTORY_PROJECT_ID,
        },
        {
          headers: {
            'Content-Type': 'application/json',
            'Authorization': `Bearer ${process.env.INFACTORY_API_KEY}`
          }
        }
      );
      
      // Step 3: Have the LLM format the response
      const formattingPrompt = `
      You are a helpful assistant. Format the following data into a natural, conversational response.
      User question: "${question}"
      Data: ${JSON.stringify(infactoryResponse.data)}
      
      Provide a conversational, easy to understand answer based strictly on this data.`;
      
      const formattingResponse = await openai.chat.completions.create({
        model: "gpt-3.5-turbo",
        messages: [{ role: "user", content: formattingPrompt }],
        temperature: 0.7,
      });
      
      responseData = {
        content: formattingResponse.choices[0].message.content,
        data: infactoryResponse.data.data,
        source: 'Infactory'
      };
    } else {
      // Step 4: If it's not a data question, let the LLM handle it
      const chatMessages = [
        { role: "system", content: "You are a helpful assistant that works with a database system. For data-specific questions, those are handled by another system. Focus on being helpful for general questions, clarifications, and conversation." },
        ...history.map(msg => ({ role: msg.role, content: msg.content })),
        { role: "user", content: question }
      ];
      
      const chatResponse = await openai.chat.completions.create({
        model: "gpt-3.5-turbo",
        messages: chatMessages,
        temperature: 0.7,
      });
      
      responseData = {
        content: chatResponse.choices[0].message.content,
        source: 'llm'
      };
    }
    
    res.json(responseData);
  } catch (error) {
    console.error('Chatbot error:', error);
    res.status(500).json({ error: 'Failed to process your request' });
  }
});

module.exports = router;

Enhancing Your Chatbot

Support for Follow-up Questions

To handle follow-up questions, maintain conversation context and reference previous responses:

// Track the last query and its parameters
let lastQuery = null;

// Example follow-up handler
function handleFollowup(question, history) {
  const lastMessage = history[history.length - 2]; // Get the last assistant message
  
  if (lastMessage && lastMessage.data && question.toLowerCase().includes('what about')) {
    // This looks like a follow-up question
    if (lastMessage.queryInfo.query_used === 'average_by_category') {
      // Extract what they're asking about
      const match = question.match(/what about (the |)([a-z0-9 ]+)/i);
      if (match) {
        const newCategory = match[2].trim();
        
        // Update the parameters from the last query
        return {
          isFollowup: true,
          queryName: lastMessage.queryInfo.query_used,
          parameters: {
            ...lastMessage.queryInfo.parameters,
            category: newCategory
          }
        };
      }
    }
  }
  
  return { isFollowup: false };
}

Adding Visualizations

Enhance your chatbot with visual representations of data:

import { useState, useRef } from 'react';
import { Bar, Line, Pie } from 'react-chartjs-2';
import { Chart as ChartJS, CategoryScale, LinearScale, BarElement, PointElement, LineElement, ArcElement, Title, Tooltip, Legend } from 'chart.js';

// Register Chart.js components
ChartJS.register(CategoryScale, LinearScale, BarElement, PointElement, LineElement, ArcElement, Title, Tooltip, Legend);

function ChatBotMessage({ message }) {
  const [showChart, setShowChart] = useState(false);
  
  // Only process data visualizations for assistant messages with data
  if (message.role !== 'assistant' || !message.data) {
    return (
      <div className={`message ${message.role}`}>
        <div className="content">{message.content}</div>
      </div>
    );
  }
  
  // Determine what type of chart to show based on the data
  const getChartType = () => {
    if (!message.data || !Array.isArray(message.data)) return null;
    
    // Example: For data that has categories and a single metric, use a bar chart
    if (message.data.length > 1 && message.data.length <= 10) {
      const firstItem = message.data[0];
      const keys = Object.keys(firstItem);
      
      if (keys.length === 2 && typeof firstItem[keys[1]] === 'number') {
        // Likely a category and a value - good for a bar chart
        return 'bar';
      } else if (keys.includes('date') || keys.includes('time') || keys.some(k => k.includes('date') || k.includes('time'))) {
        // Time series data - good for a line chart
        return 'line';
      } else if (message.data.length <= 6) {
        // Small number of categories - could be good for a pie chart
        return 'pie';
      }
    }
    
    return 'bar'; // Default to bar chart
  };
  
  const chartType = getChartType();
  
  // Prepare chart data based on the data structure
  const prepareChartData = () => {
    if (!message.data || !Array.isArray(message.data) || message.data.length === 0) return null;
    
    const firstItem = message.data[0];
    const keys = Object.keys(firstItem);
    
    // For a simple two-column data structure (category and value)
    if (keys.length === 2 && typeof firstItem[keys[1]] === 'number') {
      const categoryKey = keys[0];
      const valueKey = keys[1];
      
      return {
        labels: message.data.map(item => item[categoryKey]),
        datasets: [
          {
            label: valueKey.charAt(0).toUpperCase() + valueKey.slice(1),
            data: message.data.map(item => item[valueKey]),
            backgroundColor: 'rgba(53, 162, 235, 0.5)',
            borderColor: 'rgba(53, 162, 235, 1)',
            borderWidth: 1
          }
        ]
      };
    }
    
    // Add more data preparation logic for other data structures
    
    return null;
  };
  
  const chartData = prepareChartData();
  
  // Chart options
  const chartOptions = {
    responsive: true,
    plugins: {
      legend: {
        position: 'top',
      },
      title: {
        display: true,
        text: 'Data Visualization',
      },
    },
  };
  
  // Render the appropriate chart component based on chartType
  const renderChart = () => {
    if (!chartData) return null;
    
    switch (chartType) {
      case 'bar':
        return <Bar data={chartData} options={chartOptions} />;
      case 'line':
        return <Line data={chartData} options={chartOptions} />;
      case 'pie':
        return <Pie data={chartData} options={chartOptions} />;
      default:
        return null;
    }
  };
  
  return (
    <div className={`message ${message.role}`}>
      <div className="content">{message.content}</div>
      
      {chartType && chartData && (
        <div className="message-actions">
          <button onClick={() => setShowChart(!showChart)}>
            {showChart ? 'Hide Visualization' : 'Show Visualization'}
          </button>
          
          {showChart && (
            <div className="chart-container">
              {renderChart()}
            </div>
          )}
        </div>
      )}
      
      <div className="message-footer">
        <small>Source: Infactory data query</small>
      </div>
    </div>
  );
}

import { useState, useRef } from 'react';
import { Bar, Line, Pie } from 'react-chartjs-2';
import { Chart as ChartJS, CategoryScale, LinearScale, BarElement, PointElement, LineElement, ArcElement, Title, Tooltip, Legend } from 'chart.js';

// Register Chart.js components
ChartJS.register(CategoryScale, LinearScale, BarElement, PointElement, LineElement, ArcElement, Title, Tooltip, Legend);

function ChatBotMessage({ message }) {
  const [showChart, setShowChart] = useState(false);
  
  // Only process data visualizations for assistant messages with data
  if (message.role !== 'assistant' || !message.data) {
    return (
      <div className={`message ${message.role}`}>
        <div className="content">{message.content}</div>
      </div>
    );
  }
  
  // Determine what type of chart to show based on the data
  const getChartType = () => {
    if (!message.data || !Array.isArray(message.data)) return null;
    
    // Example: For data that has categories and a single metric, use a bar chart
    if (message.data.length > 1 && message.data.length <= 10) {
      const firstItem = message.data[0];
      const keys = Object.keys(firstItem);
      
      if (keys.length === 2 && typeof firstItem[keys[1]] === 'number') {
        // Likely a category and a value - good for a bar chart
        return 'bar';
      } else if (keys.includes('date') || keys.includes('time') || keys.some(k => k.includes('date') || k.includes('time'))) {
        // Time series data - good for a line chart
        return 'line';
      } else if (message.data.length <= 6) {
        // Small number of categories - could be good for a pie chart
        return 'pie';
      }
    }
    
    return 'bar'; // Default to bar chart
  };
  
  const chartType = getChartType();
  
  // Prepare chart data based on the data structure
  const prepareChartData = () => {
    if (!message.data || !Array.isArray(message.data) || message.data.length === 0) return null;
    
    const firstItem = message.data[0];
    const keys = Object.keys(firstItem);
    
    // For a simple two-column data structure (category and value)
    if (keys.length === 2 && typeof firstItem[keys[1]] === 'number') {
      const categoryKey = keys[0];
      const valueKey = keys[1];
      
      return {
        labels: message.data.map(item => item[categoryKey]),
        datasets: [
          {
            label: valueKey.charAt(0).toUpperCase() + valueKey.slice(1),
            data: message.data.map(item => item[valueKey]),
            backgroundColor: 'rgba(53, 162, 235, 0.5)',
            borderColor: 'rgba(53, 162, 235, 1)',
            borderWidth: 1
          }
        ]
      };
    }
    
    // Add more data preparation logic for other data structures
    
    return null;
  };
  
  const chartData = prepareChartData();
  
  // Chart options
  const chartOptions = {
    responsive: true,
    plugins: {
      legend: {
        position: 'top',
      },
      title: {
        display: true,
        text: 'Data Visualization',
      },
    },
  };
  
  // Render the appropriate chart component based on chartType
  const renderChart = () => {
    if (!chartData) return null;
    
    switch (chartType) {
      case 'bar':
        return <Bar data={chartData} options={chartOptions} />;
      case 'line':
        return <Line data={chartData} options={chartOptions} />;
      case 'pie':
        return <Pie data={chartData} options={chartOptions} />;
      default:
        return null;
    }
  };
  
  return (
    <div className={`message ${message.role}`}>
      <div className="content">{message.content}</div>
      
      {chartType && chartData && (
        <div className="message-actions">
          <button onClick={() => setShowChart(!showChart)}>
            {showChart ? 'Hide Visualization' : 'Show Visualization'}
          </button>
          
          {showChart && (
            <div className="chart-container">
              {renderChart()}
            </div>
          )}
        </div>
      )}
      
      <div className="message-footer">
        <small>Source: Infactory data query</small>
      </div>
    </div>
  );
}

Multi-turn Context Awareness

To handle complex conversations with multiple turns:

// Conversation state manager
class ConversationState {
  constructor() {
    this.topics = new Map(); // Track topics mentioned in the conversation
    this.lastQuery = null;   // Last executed query
    this.lastResponse = null; // Last response data
    this.context = {};       // Additional context like filters applied
  }
  
  updateFromMessage(message, isUser) {
    if (isUser) {
      // Process user message to extract topics, entities, etc.
      this.extractTopics(message);
    } else {
      // Update based on system response
      if (message.queryInfo) {
        this.lastQuery = message.queryInfo.query_used;
        this.lastResponse = message.data;
        
        // Update context based on the parameters used
        if (message.queryInfo.parameters) {
          Object.entries(message.queryInfo.parameters).forEach(([key, value]) => {
            this.context[key] = value;
          });
        }
      }
    }
  }
  
  extractTopics(message) {
    // Simple implementation - in production you might use NLP services
    const lowerMessage = message.toLowerCase();
    
    // Check for key topics relevant to your data domain
    const potentialTopics = ['sales', 'revenue', 'customers', 'products', 'regions'];
    
    potentialTopics.forEach(topic => {
      if (lowerMessage.includes(topic)) {
        // Increase the relevance score for this topic
        const currentScore = this.topics.get(topic) || 0;
        this.topics.set(topic, currentScore + 1);
      }
    });
  }
  
  getRelevantContext(message) {
    // Determine what context is relevant to the current message
    const context = { ...this.context };
    
    // Decay or remove irrelevant context based on the new message
    // This is highly domain-specific
    
    return context;
  }
}

Deployment Considerations

Scaling Your Chatbot

As usage grows, consider these scaling strategies:

Load Balancing

Distribute API calls across multiple backend instances

Caching

Cache common queries to reduce API calls and improve response times

Queue Processing

Use a message queue for handling high volumes of requests

Serverless Functions

Deploy as serverless functions that scale automatically with demand

Monitoring and Analytics

Track your chatbot’s performance:

Usage Metrics: Number of questions, unique users, peak usage times
Performance Metrics: Response times, API call success rates
Content Metrics: Most common questions, unanswered questions
User Satisfaction Metrics: Feedback ratings, abandonment rates

Best Practices

Clear Scope

Define what types of questions your chatbot should answer

Fallback Strategies

Have graceful responses for questions outside your query coverage

Progressive Enhancement

Start simple and add complexity as you learn user patterns

Continuous Improvement

Regularly analyze unanswerable questions to identify new queries to create

User Feedback Loop

Allow users to flag incorrect or unhelpful answers

Test with Real Users

Validate with actual users rather than just hypothetical questions

Common Challenges and Solutions

Handling Ambiguous Questions

Challenge: Users ask questions that could match multiple queries or have unclear parameters.

Solution: Implement a clarification flow where the chatbot asks follow-up questions when a question is ambiguous. For example:

if (matchingQueries.length > 1 && matchingQueries[0].confidence < 0.8) {
  return {
    needsClarification: true,
    clarificationQuestion: `I'm not sure exactly what you're asking about. Are you interested in ${matchingQueries.map(q => q.topic).join(' or ')}?`,
    matchingQueries: matchingQueries
  };
}

Maintaining Conversation Context

Challenge: Keeping track of context across multiple turns of conversation.

Solution: Implement a context management system that tracks entities, topics, and previous queries. Use this context to enhance subsequent queries.

// Update context based on the current query
function updateContext(context, query, parameters) {
  return {
    ...context,
    lastQuery: query,
    entities: {
      ...context.entities,
      ...extractEntities(parameters)
    },
    timestamp: Date.now()
  };
}

// Apply context to a new question
function applyContext(question, context) {
  // If the question includes pronouns like "it", "them", "those", 
  // try to resolve using context
  if (/\b(it|them|those|that|these)\b/i.test(question)) {
    if (context.entities.category) {
      // Replace pronouns with the actual entity from context
      return question.replace(
        /\b(it|them|those|that|these)\b/i, 
        context.entities.category
      );
    }
  }
  
  return question;
}

Unrealistic User Expectations

Challenge: Users expect the chatbot to answer any question, regardless of whether the data supports it.

Solution: Set clear expectations and provide informative responses about the chatbot’s capabilities:

function handleUnanswerable(question) {
  // Check if the question is about a topic we don't have data for
  const unsupportedTopics = ['employee salaries', 'future predictions', 'competitor data'];
  
  for (const topic of unsupportedTopics) {
    if (question.toLowerCase().includes(topic)) {
      return `I don't have information about ${topic}. I can help you with questions about our sales, customers, products, and regional performance.`;
    }
  }
  
  // General fallback
  return "I don't have enough information to answer that question. I can help you with questions about our business data such as sales figures, customer metrics, and product performance.";
}

Example Use Cases

Customer Service Bot

Answer questions about order status, shipping times, and product availability

Internal Analytics Assistant

Help employees explore business metrics without needing SQL knowledge

Sales Dashboard Companion

Provide conversational access to sales performance data

Product Recommendation Bot

Recommend products based on customer data and preferences

Next Steps

After building your chatbot, consider:

Implementing user feedback mechanisms to improve your queries
Adding authentication and authorization for internal chatbots
Exploring other integration patterns like dashboards or search interfaces
Setting up performance monitoring to ensure responsiveness

On this page

Building Chatbots with Infactory
Why Infactory for Chatbots?
Architecture Overview
Implementation Options
Simple Question-Answer Bot
Hybrid LLM + Infactory Chatbot
Enhancing Your Chatbot
Support for Follow-up Questions
Adding Visualizations
Multi-turn Context Awareness
Deployment Considerations
Scaling Your Chatbot
Monitoring and Analytics
Best Practices
Common Challenges and Solutions
Example Use Cases
Next Steps

​Building Chatbots with Infactory

​Why Infactory for Chatbots?

Hallucination

Consistency

Performance

Complexity

​Architecture Overview

​Implementation Options

​Simple Question-Answer Bot

​Hybrid LLM + Infactory Chatbot

​Enhancing Your Chatbot

​Support for Follow-up Questions

​Adding Visualizations

​Multi-turn Context Awareness

​Deployment Considerations

​Scaling Your Chatbot

Load Balancing

Caching

Queue Processing

Serverless Functions

​Monitoring and Analytics

​Best Practices

Clear Scope

Fallback Strategies

Progressive Enhancement

Continuous Improvement

User Feedback Loop

Test with Real Users

​Common Challenges and Solutions

​Example Use Cases

Customer Service Bot

Internal Analytics Assistant

Sales Dashboard Companion

Product Recommendation Bot

​Next Steps

Updates

​Building Chatbots with Infactory

​Why Infactory for Chatbots?

Hallucination

Consistency

Performance

Complexity

​Architecture Overview

​Implementation Options

​Simple Question-Answer Bot

​Hybrid LLM + Infactory Chatbot

​Enhancing Your Chatbot

​Support for Follow-up Questions

​Adding Visualizations

​Multi-turn Context Awareness

​Deployment Considerations

​Scaling Your Chatbot

Load Balancing

Caching

Queue Processing

Serverless Functions

​Monitoring and Analytics

​Best Practices

Clear Scope

Fallback Strategies

Progressive Enhancement

Continuous Improvement

User Feedback Loop

Test with Real Users

​Common Challenges and Solutions

​Example Use Cases

Customer Service Bot

Internal Analytics Assistant

Sales Dashboard Companion

Product Recommendation Bot

​Next Steps

Building Chatbots with Infactory

Why Infactory for Chatbots?

Architecture Overview

Implementation Options

Simple Question-Answer Bot

Hybrid LLM + Infactory Chatbot

Enhancing Your Chatbot

Support for Follow-up Questions

Adding Visualizations

Multi-turn Context Awareness

Deployment Considerations

Scaling Your Chatbot

Monitoring and Analytics

Best Practices

Common Challenges and Solutions

Example Use Cases

Next Steps

Building Chatbots with Infactory

Why Infactory for Chatbots?

Architecture Overview

Implementation Options

Simple Question-Answer Bot

Hybrid LLM + Infactory Chatbot

Enhancing Your Chatbot

Support for Follow-up Questions

Adding Visualizations

Multi-turn Context Awareness

Deployment Considerations

Scaling Your Chatbot

Monitoring and Analytics

Best Practices

Common Challenges and Solutions

Example Use Cases

Next Steps