config.yaml

# Dual-Model Memory System Configuration

# Database Configuration (PostgreSQL with pgvector)
database:
  host: ${DB_HOST:localhost}
  port: ${DB_PORT:5432}
  name: ${DB_NAME:memory_db}
  user: ${DB_USER:postgres}
  password: ${DB_PASSWORD:postgres}
  pool_size: 10
  max_overflow: 20

# Small Language Model (SLM) Configuration
slm:
  # Current model configuration
  current:
    name: "microsoft/phi-2"  # 2.7B parameters, ~4GB VRAM
    type: "phi"
    context_window: 2048
    quantization: "none"  # Options: "none", "8bit", "4bit" - disabled for testing
    device: "cpu"  # Changed to cpu for testing compatibility

  # Next upgrade model (for future use)
  next:
    name: "mistralai/Mistral-7B-v0.1"  # 7B parameters, ~8GB VRAM
    type: "mistral"
    context_window: 8192
    quantization: "4bit"
    device: "auto"

  # LoRA Configuration for fine-tuning
  lora:
    r: 8
    alpha: 32
    dropout: 0.1
    target_modules: ["q_proj", "v_proj", "k_proj", "o_proj"]
    bias: "none"
    task_type: "CAUSAL_LM"

  # Training Configuration
  training:
    batch_size: 4
    gradient_accumulation_steps: 4
    learning_rate: 2.0e-4
    num_epochs: 1
    warmup_steps: 100
    max_grad_norm: 1.0
    fp16: true
    logging_steps: 10
    save_steps: 100
    save_total_limit: 3

  # Retraining Schedule
  retraining:
    trigger_type: "adaptive"  # Options: "fixed", "adaptive", "manual"
    conversation_threshold: 25  # Retrain after N conversations
    time_threshold_hours: 24  # Or after N hours
    importance_threshold: 0.7  # Only train on important conversations

# Embedding Model Configuration
embeddings:
  model: "sentence-transformers/all-MiniLM-L6-v2"
  dimension: 384
  batch_size: 32
  cache_dir: "./data/embeddings_cache"

# Primary LLM Configuration
primary_llm:
  provider: "openai"  # Options: "openai", "anthropic", "local"
  model: "gpt-4o-mini"  # 125k input context window
  temperature: 0.7
  max_tokens: 4096
  context_window: 125000  # gpt-4o-mini supports 125k input tokens

  # Fallback configuration
  fallback:
    provider: "anthropic"
    model: "claude-3-sonnet-20240229"

# Memory Management
memory:
  # Working Memory (Real-time context file)
  working_memory:
    max_size: 10  # Last N conversations
    file_path: "./data/working_memory.json"
    update_frequency: "realtime"

  # Storage Configuration
  storage:
    conversations_path: "./data/conversations"
    models_path: "./data/models"
    adapters_path: "./data/models/adapters"
    checkpoints_path: "./data/models/checkpoints"
    backup_path: "./data/backups"
    backup_interval_hours: 24

  # Retention Policy
  retention:
    compress_after_days: 30  # Compress old conversations to summaries
    archive_after_days: 90   # Move to cold storage
    delete_after_days: 365   # Permanent deletion
    importance_threshold: 0.5  # Keep important conversations longer

  # Context Window Management
  context_window:
    max_tokens: 2048  # SLM context window
    overflow_strategy: "chunking"  # Options: "chunking", "summarization", "hybrid"
    chunk_size: 1024
    chunk_overlap: 128
    priority_scoring: true  # Prioritize important chunks

# Tool Call Interface
tools:
  enabled: true
  timeout_seconds: 30
  max_retries: 3

  # Available tools for Primary LLM
  available:
    - name: "memory_search"
      description: "Search memory for relevant conversations and facts"
      parameters:
        - query: "string"
        - query_type: "semantic|temporal|factual"
        - limit: "integer"

    - name: "memory_store"
      description: "Store new information in memory"
      parameters:
        - content: "string"
        - importance: "float"
        - tags: "array"

    - name: "get_context"
      description: "Get recent conversation context"
      parameters:
        - num_recent: "integer"

    - name: "summarize_memory"
      description: "Get summary of stored memories"
      parameters:
        - time_range: "string"
        - topics: "array"

# API Configuration
api:
  host: "0.0.0.0"
  port: 8000
  workers: 4
  reload: false
  log_level: "info"
  cors_origins: ["*"]
  rate_limit:
    enabled: true
    requests_per_minute: 60

# Monitoring & Logging
monitoring:
  enabled: true
  metrics_port: 9090
  log_file: "./logs/memory_system.log"
  log_level: "INFO"
  log_format: "json"

  # Metrics to track
  metrics:
    - "conversation_count"
    - "memory_query_latency"
    - "training_time"
    - "model_size"
    - "database_size"
    - "cache_hit_rate"

# Docker Configuration
docker:
  volumes:
    - "./data:/app/data"
    - "./logs:/app/logs"
    - "./models:/app/models"

  environment:
    PYTHONUNBUFFERED: "1"
    CUDA_VISIBLE_DEVICES: "0"
    TRANSFORMERS_CACHE: "/app/models/cache"

# Feature Flags
features:
  enable_compression: true
  enable_summarization: true
  enable_user_preferences: true  # Support # commands for preferences
  enable_auto_tagging: true
  enable_fact_extraction: true
  enable_multi_user: false  # Future feature
  enable_distributed_training: false  # Future feature

# Performance Tuning
performance:
  use_gpu: true
  mixed_precision: true
  gradient_checkpointing: true
  optimize_embeddings: true
  cache_embeddings: true
  async_training: true
  parallel_queries: true