trading-platform/apps/llm-agent

LLM Agent Service

AI-powered trading copilot with local GPU support for OrbiQuant IA Trading Platform.

Overview

This service provides an intelligent trading agent that runs locally on GPU using Ollama, with support for:

• Local LLM Inference: Run Llama 3, Mistral, and other models on your GPU
• Trading Analysis: Technical analysis, market sentiment, AMD phase identification
• Function Calling (Tools): Get real-time data, ML signals, portfolio info
• Educational Support: Explain concepts, recommend learning paths
• Risk Management: Position sizing, stop loss recommendations
• Streaming Responses: Real-time chat with Server-Sent Events

Why Local LLM?

• Privacy: Your trading conversations stay on your hardware
• Cost: No API costs after initial setup
• Speed: Low latency with local GPU
• Customization: Full control over model and prompts
• Always Available: No internet dependency for inference

Technology Stack

• Python: 3.11+
• LLM Provider: Ollama (local GPU) + Claude/OpenAI fallback
• Models: Llama 3 8B (recommended), Mistral 7B, Mixtral 8x7b
• API Framework: FastAPI with SSE streaming
• Tools System: Custom function calling for trading operations
• Database: PostgreSQL (asyncpg)
• Cache: Redis
• Testing: pytest, pytest-asyncio

Setup

Prerequisites

• GPU: NVIDIA RTX 5060 Ti (16GB VRAM) or better
• RAM: 16GB minimum, 32GB recommended
• OS: Linux (Ubuntu 20.04+) or WSL2
• Software:
  • Docker with NVIDIA Container Toolkit
  • Python 3.11+
  • Miniconda or Anaconda

Quick Start

1. Start Ollama:

cd /home/isem/workspace/projects/trading-platform/apps/llm-agent

# Start Ollama with GPU support
docker-compose -f docker-compose.ollama.yml up -d

# Pull Llama 3 8B model (recommended for 16GB VRAM)
docker exec orbiquant-ollama ollama pull llama3:8b

2. Configure Service:

# Copy environment file
cp .env.example .env

# Edit configuration (LLM_PROVIDER=ollama is default)
nano .env

3. Install Dependencies:

# Create conda environment
conda env create -f environment.yml
conda activate orbiquant-llm-agent

# Or install with pip
pip install -r requirements.txt

4. Start the Service:

# Development mode with hot-reload
uvicorn src.main:app --reload --host 0.0.0.0 --port 8003

5. Test it:

Open in browser: http://localhost:8003/docs

# Or via curl
curl http://localhost:8003/api/v1/health

Detailed Setup

See DEPLOYMENT.md for:

• Complete installation guide
• GPU configuration
• Model selection
• Performance tuning
• Troubleshooting

Project Structure

llm-agent/
├── src/
│   ├── main.py                    # FastAPI application
│   ├── config.py                  # Configuration management
│   ├── core/                      # Core LLM functionality
│   │   ├── llm_client.py          # Ollama/Claude client
│   │   ├── prompt_manager.py      # System prompts
│   │   └── context_manager.py     # Conversation context
│   ├── tools/                     # Trading tools (function calling)
│   │   ├── base.py                # Tool base classes
│   │   ├── signals.py             # ML signals & analysis
│   │   ├── portfolio.py           # Portfolio management
│   │   ├── trading.py             # Trading execution
│   │   └── education.py           # Educational tools
│   ├── prompts/                   # System prompts
│   │   ├── system.txt             # Main trading copilot prompt
│   │   ├── analysis.txt           # Analysis template
│   │   └── strategy.txt           # Strategy template
│   ├── api/                       # API routes
│   │   └── routes.py              # All endpoints
│   ├── models/                    # Pydantic models
│   ├── services/                  # Business logic
│   └── repositories/              # Data access layer
├── tests/
├── docker-compose.ollama.yml      # Ollama GPU setup
├── DEPLOYMENT.md                  # Detailed deployment guide
├── requirements.txt
├── environment.yml
└── .env.example

API Endpoints

Core Endpoints

• GET / - Service info and health
• GET /api/v1/health - Detailed health check with LLM status
• GET /api/v1/models - List available LLM models

Chat & Analysis

• POST /api/v1/chat - Interactive chat (supports streaming)
• POST /api/v1/analyze - Comprehensive symbol analysis
• POST /api/v1/strategy - Generate trading strategy
• POST /api/v1/explain - Explain trading concepts

Tools & Context

• GET /api/v1/tools - List available tools for user plan
• DELETE /api/v1/context/{user_id}/{conversation_id} - Clear conversation

See interactive docs at: http://localhost:8003/docs

Development

Code Quality

# Format code
black src/
isort src/

# Lint
flake8 src/

# Type checking
mypy src/

Testing

# Run all tests
pytest

# With coverage
pytest --cov=src --cov-report=html

# Specific tests
pytest tests/unit/

Trading Tools (Function Calling)

The agent has access to these tools:

Market Data (Free)

• get_analysis - Current price, volume, 24h change
• get_news - Recent news with sentiment analysis
• calculate_position_size - Risk-based position sizing

ML & Signals (Pro/Premium)

• get_signal - ML predictions with entry/exit levels
• check_portfolio - Portfolio overview and P&L
• get_positions - Detailed position information
• get_trade_history - Historical trades with metrics

Trading (Pro/Premium)

• execute_trade - Execute paper trading orders
• set_alert - Create price alerts

Education (Free)

• explain_concept - Explain trading terms (RSI, MACD, AMD, etc.)
• get_course_info - Recommend learning resources

Tools are automatically filtered based on user subscription plan (Free/Pro/Premium).

System Prompt & Trading Philosophy

The agent is trained with a comprehensive system prompt that includes:

• AMD Framework: Identifies Accumulation, Manipulation, and Distribution phases
• Risk Management: Always prioritizes proper position sizing and stop losses
• Educational Approach: Explains the "why" behind every recommendation
• Multi-timeframe Analysis: Considers multiple timeframes for context
• Data-Driven: Uses tools to fetch real data, never invents prices

The agent will:

• ✅ Provide educational analysis with clear risk management
• ✅ Explain concepts in simple terms for all levels
• ✅ Use real market data via tools
• ✅ Warn against risky behavior
• ❌ NEVER give financial advice or guarantee returns
• ❌ NEVER invent market data or prices

Architecture

Built following SOLID principles:

• LLM Client Abstraction: Unified interface for Ollama/Claude/OpenAI
• Tool Registry: Dynamic tool loading with permission checking
• Context Manager: Maintains conversation history efficiently
• Prompt Manager: Centralized prompt templates
• Streaming Support: Real-time responses with SSE

Configuration

Key environment variables:

# LLM Provider
LLM_PROVIDER=ollama              # ollama, claude, or openai
OLLAMA_BASE_URL=http://localhost:11434
LLM_MODEL=llama3:8b

# Service URLs
BACKEND_URL=http://localhost:8000
DATA_SERVICE_URL=http://localhost:8001
ML_ENGINE_URL=http://localhost:8002

# Optional: Claude fallback
# ANTHROPIC_API_KEY=sk-ant-xxx

# Database & Cache
DATABASE_URL=postgresql://...
REDIS_URL=redis://localhost:6379/0

Model Recommendations

For RTX 5060 Ti (16GB VRAM):

Model	Size	Speed	Quality	Best For
llama3:8b	4.7GB	⚡⚡⚡	⭐⭐⭐⭐	Recommended - Best balance
mistral:7b	4.1GB	⚡⚡⚡⚡	⭐⭐⭐	Fast responses, good quality
llama3:70b	40GB+	⚡	⭐⭐⭐⭐⭐	Requires 40GB+ VRAM
mixtral:8x7b	26GB	⚡⚡	⭐⭐⭐⭐	Requires 32GB+ VRAM

Recommendation: Start with llama3:8b for your hardware.

Documentation

• DEPLOYMENT.md - Complete deployment guide
• API Documentation - Interactive API docs
• Specification Docs - Technical specifications

License

Proprietary - OrbiQuant IA Trading Platform