AgentFlow

AI-powered customer service conversation management, automation, and analytics platform

Overview

AgentFlow is a comprehensive AI-powered customer service platform that provides real-time conversation management, intelligent intent detection, automated FAQ responses, and model fine-tuning capabilities. Built on Amazon Bedrock AI, it extracts intent classifications, identifies key parameters, manages conversation flows, and provides interactive visualizations to understand and optimize customer service operations.

Key Features

• AI-Powered Classification: Uses LLM-based intent parsing on Amazon Bedrock to intelligently classify conversation intents
• Interactive Visualizations: Sankey diagrams showing conversation flow patterns
• Parameter Extraction: Automatically identifies customer-provided and agent-needed parameters
• Visual Schema Editor: Drag-and-drop interface for managing intent hierarchies
• Flow Analytics: Detailed statistics and insights for each conversation flow
• Real-time Processing: Automatic classification triggered by S3 uploads
• Centralized Configuration System: Dynamic intent management with web-based configuration interface
• Redis Session Management: Real-time conversation state tracking and caching
• FAQ Management: Intelligent FAQ responses with trigger phrase matching
• Model Fine-tuning: HuggingFace integration for custom model training
• Conversation Flow Orchestration: Advanced context switching and parameter collection

Architecture

Backend (AWS CDK)

• S3 Bucket: Stores transcript files and triggers processing
• Lambda Functions: AI classification, conversation management, and API handling
• DynamoDB Tables: Stores schemas, results, extracted parameters, and FAQ sources
• API Gateway: RESTful API for frontend integration with conversation endpoints
• Amazon Bedrock: AI model for intent classification and parameter extraction
• ElastiCache Redis: Session management, caching, and real-time training status tracking
• VPC Infrastructure: Secure networking for Lambda functions and Redis access

Frontend (React + TypeScript)

• Visual Schema Editor: Drag-and-drop intent management
• Sankey Diagrams: Interactive conversation flow visualization
• Flow Analytics: Detailed statistics and parameter analysis
• Real-time Updates: Live data synchronization
• Configuration Manager: Web-based interface for managing intents, categories, and settings
• FAQ Manager: Interactive FAQ response management and testing
• HuggingFace Fine-tuner: Model training interface with real-time progress tracking
• Conversation Tester: Interactive testing of conversation flows and intent detection

Screenshots

Note: The following screenshots showcase the platform's core capabilities including transcript analysis, conversation management, and model training interfaces.

Main Dashboard

The main dashboard showing schema management, and AI suggestions

Visual Schema Editor

Drag-and-drop interface for managing hierarchical intent structures with multi-level support

Sankey Flow Visualization

Interactive Sankey diagram showing conversation flow patterns and frequencies

Flow Statistics

Detailed analytics showing top parameters, conversation counts, and insights for each flow

Model Training Interface

HuggingFace fine-tuning interface with real-time progress tracking and model management

Conversation Flow Management

Interactive conversation testing and flow orchestration with context switching

Quick Start

Prerequisites

• AWS Account with Bedrock access
• Node.js 18+ and npm
• AWS CLI configured

1. Clone and Setup

git clone <repository-url>
cd bedrock-transcript-classifier

2. Deploy Backend

cd cdk
npm install
npm run build
npx cdk deploy

3. Start Frontend

cd ../frontend/frontend
npm install
npm run dev

4. Configure API Endpoint

• Open the frontend at http://localhost:5173
• Click the settings icon in the header
• Enter your API Gateway URL from the CDK deployment

API Endpoints

Public Endpoints (No Authentication Required)

• /intents - Get intent schemas for transcript analysis frontend
• /schemas - Manage intent schema definitions
• /results - Access transcript analysis results
• /transcripts - Transcript management operations

Conversation Endpoints (Requires Authentication)

• /conversation - Process conversation messages
• /conversation/config/* - Intent and configuration management
• /conversation/finetune/* - Model fine-tuning operations
• /conversation/testing/* - Model testing and evaluation
• /conversation/history/* - Conversation session management

Further Documentation

• CDK Backend Setup - Complete backend deployment and configuration
• Frontend Development - Frontend development and component architecture

Core Features

Intent Classification

The system automatically classifies conversations into hierarchical intent structures using both rule-based and AI-powered approaches

Parameter Extraction

Automatically identifies two types of parameters:

Input Parameters (Customer Provides):

• Account numbers, passwords, addresses
• Personal information for authentication
• Request details and preferences

Output Parameters (Agent Needs):

• Backend system data requirements
• Information needed to fulfill requests
• System queries and lookups

Centralized Configuration System

• Dynamic Intent Management: Web-based interface for creating, updating, and deleting intents
• Category Organization: Hierarchical organization of intents by business domain
• Response Templates: Configurable response patterns for each intent
• Parameter Schemas: Structured parameter definitions with validation rules
• FAQ Integration: Intelligent FAQ responses with confidence scoring

Interactive Visualizations

• Sankey Diagrams: Visual representation of conversation flows
• Click-to-Explore: Click any node to see detailed statistics
• Flow Analytics: Comprehensive metrics for each intent path
• Parameter Analysis: Top input/output parameters for each flow

Visual Schema Editor

• Drag-and-Drop: Intuitive interface for managing intents
• Multi-Level Support: Handle unlimited sub-intent levels
• JSON Sync: Bidirectional synchronization with raw JSON
• Real-time Validation: Instant feedback on schema structure

Conversation Management

• Session Tracking: Redis-powered conversation state management
• Context Switching: Intelligent context transitions based on user intent
• Parameter Collection: Progressive parameter gathering with validation
• Response Generation: Dynamic response creation using configured templates

Data Flow

1. Transcript Upload & Real-time Conversation

• Transcript Processing: User uploads a transcript file to the S3 bucket, triggering automatic classification
• Real-time Conversations: Users can interact directly with the system for immediate intent detection and response generation
• Session Management: Each conversation is tracked with Redis for context persistence and state management

2. AI Classification

• The Lambda function sends the transcript text to Amazon Bedrock
• Bedrock analyzes the conversation and determines the intent category and subcategory
• The system stores the classification result with a confidence score

3. Parameter Extraction

• The system makes two separate calls to Bedrock for parameter extraction:
  • Input Parameters: Identifies information the customer provides (account numbers, passwords, addresses, etc.)
  • Output Parameters: Identifies information the agent needs from backend systems (account balances, service status, etc.)

4. Data Storage

• Classification results are stored in the ParsedResultsTable
• Input parameters are stored in the InputParametersTable
• Output parameters are stored in the OutputParametersTable
• Each transcript gets a unique ID that links all related data

5. Frontend Access

• The React frontend fetches data from the API Gateway
• API Gateway routes requests to the API Lambda function
• The Lambda function queries DynamoDB tables and returns the data
• The frontend displays the data in visualizations (Sankey diagrams, statistics, etc.)

6. Configuration Management

• Use the Configuration Manager in the frontend to manage intents, categories, and settings
• All configuration changes are immediately reflected in the conversation system
• FAQ responses can be managed and tested through the FAQ Manager interface
• Intent schemas are automatically synchronized between the configuration system and transcript analysis

Upcoming Features

• Handling multiple intents per transcript
• Cleaning up the schema editor
• Proper data obfuscation for sensitive data
• Support for more cloud providers (Azure, GCP, etc.)
• Cleaner onboarding experience
• Better error handling
• Better testing (unit tests, integration tests, etc.)

IN-PROGRESS

Currently In-Progress Features

RAG (Retrieval-Augmented Generation)

• Status: Partially implemented but disabled
• Issue: OpenSearch integration incomplete, vector store not fully operational
• Current State: Placeholder endpoints exist but functionality is limited
• Next Steps: Complete OpenSearch domain setup and vector embedding pipeline

LangChain Integration

• Status: Simplified implementation in place
• Issue: Full LangChain framework not integrated due to Lambda size constraints
• Current State: Custom conversation orchestration using simplified patterns
• Next Steps: Optimize Lambda packaging or implement lightweight LangChain subset

HuggingFace + LoRA Fine-tuning

• Status: Infrastructure deployed, training simulation active
• Issue: Real model training not implemented due to Lambda execution time limits
• Current State: Simulated training process with Redis progress tracking
• Next Steps: Implement actual HuggingFace API integration or move to ECS/Fargate for training

Advanced Model Management

• Status: Basic model storage and testing implemented
• Issue: Model versioning, A/B testing, and production deployment incomplete
• Current State: Simple model storage with basic inference testing
• Next Steps: Implement model registry, versioning, and deployment pipeline

Real-time Training Pipeline

• Status: Architecture designed, implementation partial
• Issue: Continuous learning from conversation feedback not operational
• Current State: Manual training job initiation and monitoring
• Next Steps: Implement automated feedback collection and incremental training

License

This project is licensed under the MIT License - see the LICENSE file for details.