AI-Powered Trading System with Local LLM

A sophisticated trading system leveraging local LLM deployment through Ollama, distributed computing with Apache Spark, vector-based episodic memory using ChromaDB, and advanced RAG (Retrieval Augmented Generation) capabilities to make intelligent trading decisions.

💡 Project status: In progress

🚀 Key Features

1. Local LLM Integration

   • Custom finance-tuned LLM model served through Ollama
   • Optimized for trading analysis and decision making
   • No dependency on external API services
   • Customizable model parameters and system prompts

2. Real-time Processing

   • Apache Spark distributed computing
   • Kafka stream processing
   • Real-time technical analysis
   • Multi-symbol monitoring

3. Advanced Memory System

   • ChromaDB vector store for episodic memory
   • Semantic search capabilities
   • Memory importance scoring
   • Automatic memory consolidation
   • Context-aware retrieval

4. Contextual RAG System

   • Hybrid retrieval combining BM25 and dense embeddings
   • Cross-encoder reranking for improved relevance
   • Market context-aware document retrieval
   • Automatic index maintenance and optimization
   • Time-window based filtering
   • Configurable relevance thresholds

5. Intelligent Analysis

   • Multi-source research aggregation
   • Technical indicator analysis
   • Market context integration
   • Sentiment analysis
   • Historical pattern recognition
   • Risk assessment

🛠️ Technical Stack

• LLM: Custom finance-tuned model via Ollama
• Vector Store: ChromaDB with DuckDB
• RAG System: Hybrid BM25 + Dense Retrieval
• Reranking: Cross-encoder (MiniLM)
• Distributed Computing: Apache Spark
• Stream Processing: Kafka
• Market Data: Alpaca API
• Embeddings: OpenAI Ada 002
• Text Processing: NLTK, Transformers

🔄 Architecture

┌─────────────────┐    ┌──────────────────┐    ┌────────────────┐
│  Market Data    │───▶│  Spark Streaming │───▶│  Technical     │
│  Stream (Alpaca)│    │  Processing      │    │  Analysis      │
└─────────────────┘    └──────────────────┘    └────────────────┘
                                                       │
┌─────────────────┐    ┌──────────────────┐            ▼
│  Research       │───▶│  Local LLM       │    ┌────────────────┐
│  Aggregation    │    │  & RAG System    │◀───│  Trading Agent │
└─────────────────┘    └──────────────────┘    └────────────────┘
                              ▲                        │
┌─────────────────┐    ┌──────┴───────────┐            ▼
│  ChromaDB       │◀──▶│  Memory System   │    ┌────────────────┐
│  Vector Store   │    │  & BM25 Index    │◀───│  Decision      │
└─────────────────┘    └──────────────────┘    │  Engine        │
                                               └────────────────┘

💡 System Components

RAG System

• Hybrid Retrieval: Combines BM25 keyword search with dense embeddings
• Reranking: Uses cross-encoder for improved result relevance
• Context Integration: Incorporates market conditions and temporal context
• Automatic Maintenance: Regular index updates and optimization
• Configurable Parameters: Adjustable relevance thresholds and result limits

Memory Management

• Vector Storage: Efficient storage and retrieval of trading decisions
• Contextual Search: Market-aware memory retrieval
• Importance Scoring: Automatic scoring of trading decisions
• Memory Consolidation: Periodic synthesis of related memories
• Temporal Awareness: Time-based filtering and relevance decay

Analysis Pipeline

• Multi-source Research: Aggregates data from various sources
• Technical Analysis: Real-time indicator calculation
• Market Context: Dynamic market condition assessment
• Risk Analysis: Comprehensive risk factor evaluation
• Pattern Recognition: Historical pattern matching and analysis

⚙️ Configuration

The system uses environment variables for configuration. Key RAG-related settings:

RAG_INDEX_DIR=./rag_indices
RAG_MIN_RELEVANCE_SCORE=0.5
RAG_UPDATE_INTERVAL_HOURS=4
RAG_MAX_RESULTS=5

🤝 Contributing

Contributions are welcome! Please check the Contributing Guide for guidelines.

📝 License

MIT License - see the LICENSE file for details