Conversational Analysis Engine (CAE)

Optimize LLM responses using search algorithms - A production-ready FastAPI backend and MCP server that leverages Monte Carlo Tree Search (MCTS) to intelligently explore and evaluate conversation paths for enhanced response quality.

Table of Contents

• Quick Start (Docker Compose)
• Overview
• Architecture
• Service Modes & Usage
• Claude Desktop Integration
• Manual Setup
• Development
• Testing
• Contributing
• Roadmap
• License

Quick Start (Docker Compose)

Recommended Installation - Automatically provisions Redis, PostgreSQL/PGVector, and Prometheus:

Prerequisites

• Docker ≥ 24.0
• Docker Compose v2

One-Command Setup

git clone https://github.com/yourusername/ConversationalAnalysisEngine
cd ConversationalAnalysisEngine
docker compose up --build

What You Get

• API Backend: http://localhost:8000 (health checks, conditional metrics)
• MCP Server: http://localhost:8001/mcp/v1 (conversation analysis)
• Redis Cache: localhost:6379
• PostgreSQL/PGVector: localhost:5432

Optional Monitoring (Add --profile monitoring)

docker compose --profile monitoring up --build

• Prometheus Metrics: http://localhost:9090
• Grafana Dashboard: http://localhost:3000 (admin/admin)

Shutdown

docker compose down -v

Overview

The Conversational Analysis Engine (CAE) enhances LLM response optimization by applying advanced search algorithms to conversation paths. Instead of generating single responses, CAE uses Monte Carlo Tree Search (MCTS) to:

1. Generate multiple response branches for any conversation context
2. Simulate conversation continuations to predict outcomes
3. Score paths based on goal-specific metrics (emotional intelligence, persuasiveness, helpfulness)
4. Select the optimal response through intelligent exploration

Dual Architecture:

• FastAPI Backend: Production-ready API with health checks, metrics, and monitoring
• MCP Server: Monte Carlo Tree Search optimization via Model Context Protocol

Architecture

[Claude Desktop/Code] → [MCP Server :8001] ↔ [MCTS Algorithm] ↔ [Redis Cache]
[MCP Clients]         →                                      ↓
                                                    [Response Generator]
[Health Checks]       → [API Backend :8000] ← metrics ← [Prometheus :9090]
[Monitoring]          →                              ↓
                                            [PostgreSQL/PGVector :5432]

Service Architecture:

• MCP Server (Port 8001): Conversation analysis via MCTS algorithm
• API Backend (Port 8000): Health checks, conditional metrics endpoints
• Redis (Port 6379): Required - Conversation storage and caching
• PostgreSQL/PGVector (Port 5432): Required - Conversation storage
• Prometheus (Port 9090): Optional - Metrics collection (monitoring profile)
• Grafana (Port 3000): Optional - Metrics dashboard (monitoring profile)

Core Components

• MCTS Algorithm (app/services/mcts/): Monte Carlo Tree Search implementation with UCB1 exploration
• Response Generator: Creates diverse response branches using LLM variations
• Conversation Simulator: Predicts user reactions and conversation continuations
• Conversation Scorer: Evaluates path quality based on customizable metrics
• Semantic Cache: Redis-based caching with embedding similarity for performance optimization
• Metrics Collection: Prometheus metrics for production monitoring

Service Modes & Usage

A. MCP Server (Recommended)

Docker (Recommended):

docker compose up mcp

Local Development:

poetry run python servers/mcp/mcts_analysis_server.py --transport http --port 8001

Features:

• ✅ MCTS-Powered Conversation Analysis: Multi-branch exploration with intelligent search
• ✅ Goal-Oriented Optimization: Customize for empathy, persuasion, problem-solving
• ✅ Configurable Parameters: Branch count, simulation depth, exploration constants
• ✅ Real-time Processing: Efficient async processing with resource management

B. API Backend (Health & Monitoring)

Docker:

docker compose up api

Local Development:

poetry run python -m app.main

Features:

• ✅ Health checks at GET /health
• ✅ Conditional Prometheus metrics at GET /metrics (when enabled)
• ✅ Service monitoring and logging

⚠️ Deprecated: REST Analysis Endpoint

WARNING: The POST /api/v1/analyze endpoint is deprecated and returns HTTP 410.

Migration Path: Use the MCP server for all conversation analysis:

# ❌ Deprecated - DO NOT USE
response = httpx.post("http://localhost:8000/api/v1/analyze", ...)

# ✅ Use MCP Server instead
from mcp import Client
client = Client("http://localhost:8001/mcp/v1")
result = await client.call_tool("analyze_conversation", ...)

MCP Tool Usage

The server exposes the analyze_conversation tool with the following signature:

// TypeScript/JavaScript MCP Client Example
import { Client } from '@modelcontextprotocol/sdk/client/index.js';

const client = new Client({
    name: "cae-client",
    version: "1.0.0"
});

const result = await client.callTool("analyze_conversation", {
    conversation_goal: "help user feel better about their situation", 
    messages: [
        {role: "user", content: "I failed my exam and feel terrible"},
        {role: "assistant", content: "I'm sorry to hear about your exam."}
    ],
    num_branches: 3,
    simulation_depth: 2,
    mcts_iterations: 10
});

console.log("Optimized response:", result.selected_response);
console.log("Analysis:", result.analysis);

Claude Desktop Integration

To use CAE with Claude Desktop, add the MCP server to your configuration:

1. Open Claude Desktop configuration:

   • macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
   • Windows: %APPDATA%/Claude/claude_desktop_config.json

2. Add CAE MCP server configuration:

{
  "mcpServers": {
    "conversational-analysis-engine": {
      "command": "docker",
      "args": [
        "compose", "-f", "/path/to/ConversationalAnalysisEngine/docker-compose.yml", 
        "up", "mcp", "--build"
      ],
      "env": {
        "LLM_API_KEY": "your_openai_api_key"
      }
    }
  }
}

3. Restart Claude Desktop to load the MCP server

4. Use in conversations:

I need to respond to a difficult customer complaint. Can you use the MCTS analysis to help me find the best response?

Goal: Maintain customer relationship while addressing concerns
Current conversation: [customer complaint details]

Claude Code Integration

For Claude Code users, configure the MCP server in your settings:

{
  "mcp": {
    "servers": {
      "cae": {
        "command": "docker",
        "args": [
          "compose", "-f", "/path/to/ConversationalAnalysisEngine/docker-compose.yml",
          "up", "mcp", "--build"
        ],
        "env": {
          "LLM_API_KEY": "your_openai_api_key"
        }
      }
    }
  }
}

Manual Setup

For advanced users who prefer manual installation:

Prerequisites

• Python 3.12+
• Poetry (package manager)
• Redis (required for caching)
• PostgreSQL with PGVector (required for conversation storage)

Installation

# Clone the repository
git clone https://github.com/yourusername/ConversationalAnalysisEngine
cd ConversationalAnalysisEngine

# Install dependencies with Poetry
poetry install

# Set up environment variables
cp .env.example .env
# Edit .env with your configuration

Environment Configuration

Minimal Setup - Only one environment variable required:

# REQUIRED: LLM Configuration
LLM_API_KEY=your_openai_api_key

Full Setup - All optional configuration with smart defaults:

# REQUIRED: LLM Configuration
LLM_API_KEY=your_openai_api_key

# LLM Configuration (optional - smart defaults)
LLM_API_BASE_URL=https://api.openai.com/v1  # Default
LLM_MODEL_NAME=o3-mini                      # Default

# OPTIONAL: Embedding Configuration (enables semantic caching when present)
EMBEDDING_MODEL_API_KEY=your_openai_api_key      # Optional
EMBEDDING_MODEL_BASE_URL=https://api.openai.com/v1  # Default
EMBEDDING_MODEL_NAME=text-embedding-3-large      # Default

# Feature Toggles (optional)
DISABLE_PROMETHEUS_METRICS=false  # Default: metrics enabled

# Database Configuration (Docker Compose defaults)
DB_HOST=postgres           # Default for Docker
DB_PORT=5432              # Default
DB_NAME=conversation_analysis  # Default
DB_USER=cae_user          # Default
DB_SECRET=cae_password    # Default

# Redis Configuration (Docker Compose defaults)
REDIS_HOST=redis          # Default for Docker
REDIS_PORT=6379           # Default

# Application Settings (optional)
LOG_LEVEL=INFO            # Default
LLM_TIMEOUT_SECONDS=600   # Default

Alternative Providers (e.g., OpenRouter, Groq):

# OpenRouter Example
LLM_API_KEY=your_openrouter_api_key
LLM_API_BASE_URL=https://openrouter.ai/api/v1
LLM_MODEL_NAME=anthropic/claude-3-sonnet

# Groq Example  
LLM_API_KEY=your_groq_api_key
LLM_API_BASE_URL=https://api.groq.com/openai/v1
LLM_MODEL_NAME=llama-3.1-8b-instant

# For semantic caching with different embedding provider
EMBEDDING_MODEL_API_KEY=your_embedding_provider_key
EMBEDDING_MODEL_BASE_URL=https://api.your-provider.com/v1
EMBEDDING_MODEL_NAME=your-embedding-model

Infrastructure Setup

# Start only infrastructure with Docker Compose
docker compose -f compose.infrastructure.yml up

# Or start services manually:
redis-server
# Configure PostgreSQL with PGVector extension

MCTS Configuration

Customize search behavior through configuration:

# High-quality, slower analysis
config = {
    "num_branches": 8,           # More initial branches
    "mcts_iterations": 20,       # More iterations
    "simulation_depth": 4,       # Deeper simulations
    "exploration_constant": 1.0  # Balanced exploration
}

# Fast, real-time analysis  
config = {
    "num_branches": 3,
    "mcts_iterations": 5,
    "simulation_depth": 2,
    "exploration_constant": 2.0  # More exploration
}

Development

Code Quality

This project maintains high code quality standards:

# Linting and formatting
poetry run ruff format .
poetry run ruff check .

# Type checking
poetry run mypy app/

# Run all quality checks
make quality-check

Development Servers

# Start API backend with hot reload
poetry run uvicorn app.main:app --reload --port 8000

# Start MCP server with debug logging
poetry run python servers/mcp/mcts_analysis_server.py --log-level DEBUG

# Start both with Docker Compose
docker compose up --build

Testing

Comprehensive test suite with 95%+ coverage requirement:

# Run all tests
poetry run pytest

# With coverage report
poetry run pytest --cov=app --cov-report=html --cov-report=term

# Run specific test categories
poetry run pytest tests/unit/           # Unit tests
poetry run pytest tests/integration/   # Integration tests
poetry run pytest tests/e2e/          # End-to-end tests

# Performance tests
poetry run pytest tests/performance/ -v

Test Structure:

• Unit Tests: Individual component testing with mocks
• Integration Tests: Service interaction testing
• E2E Tests: Full workflow testing via API/MCP
• Performance Tests: Load and latency testing

Contributing

Open-source contributions are welcome! Please follow these guidelines:

Development Setup

1. Fork the repository on GitHub
2. Clone your fork: git clone https://github.com/yourusername/ConversationalAnalysisEngine
3. Install dependencies: poetry install
4. Create a feature branch using the naming convention:

# Branch naming format: feature/<feature-abbrev>-<issue-num>-<tag-line>
git checkout -b feature/CRITICAL-1-CORS-fix
git checkout -b feature/PERF-23-redis-optimization
git checkout -b feature/MCTS-45-branching-strategy

Code Standards

• Linter: Use ruff for code formatting and linting
• Test Coverage: Maintain ≥95% test coverage for all new code
• Type Hints: All functions must have proper type annotations
• Documentation: Update docstrings and README for new features

Pull Request Process

1. Status Checks: Ensure all CI checks pass (tests, linting, coverage)
2. PR Approval: At least one approval required from maintainer
3. Branch Protection: Feature branches must be up-to-date with main
4. Documentation: Update relevant documentation for new features

Commit Guidelines

# Good commit messages
git commit -m "feat: add semantic caching for MCTS nodes"
git commit -m "fix: handle timeout errors in conversation simulation"
git commit -m "docs: update API examples in README"

Roadmap

Current Focus: Performance & Scalability

• 🚧 Unified LLM Evaluation: Combine multiple LLM calls into single requests (66% cost reduction)
• 🚧 Advanced Semantic Caching: Embedding-based cache with similarity detection
• 🚧 Resource-Aware Scheduling: Dynamic resource allocation and request prioritization

Near-Term: Domain Abstraction (Coming Soon)

• 🔮 Domain-Agnostic MCTS: Generalized search framework for various optimization tasks beyond conversation
• 🔮 Alternative Search Algorithms: Beam search, A* with heuristics, hybrid approaches
• 🔮 Multi-Objective Optimization: Simultaneous optimization for multiple conversation goals

Future Enhancements: Advanced Search Mechanisms

• 🔮 Reinforcement Learning Integration: Learning-based path selection improvements
• 🔮 Distributed Processing: Horizontal scaling with work queues
• 🔮 Model Cascading: Use smaller models for simulation, larger for final generation
• 🔮 Advanced Analytics: Conversation pattern analysis and success prediction

License

This project is licensed under the MIT License.

---

Acknowledgments

• FastMCP for excellent MCP server framework
• OpenAI for OpenAI API SPec
• Anthropic for MCP specification and Claude integration

Support

• Documentation: Detailed guides in /docs
• Issues: Report bugs and feature requests via GitHub Issues

---

Built by Manav Pandey for the AI community - enabling smarter conversations through algorithmic optimization.