Timmy-time-dashboard/config/providers.yaml

# Cascade LLM Router Configuration
# Providers are tried in priority order (1 = highest)
# On failure, automatically falls back to next provider

cascade:
  # Timeout settings
  timeout_seconds: 30
  
  # Retry settings
  max_retries_per_provider: 2
  retry_delay_seconds: 1
  
  # Circuit breaker settings
  circuit_breaker:
    failure_threshold: 5        # Open circuit after 5 failures
    recovery_timeout: 60        # Try again after 60 seconds
    half_open_max_calls: 2      # Allow 2 test calls when half-open

providers:
  # Primary: Local Ollama (always try first for sovereignty)
  - name: ollama-local
    type: ollama
    enabled: true
    priority: 1
    url: "http://localhost:11434"
    models:
      # Text + Tools models
      - name: qwen3:30b
        default: true
        context_window: 128000
        # Note: actual context is capped by OLLAMA_NUM_CTX (default 4096) to save RAM
        capabilities: [text, tools, json, streaming]
      - name: llama3.1:8b-instruct
        context_window: 128000
        capabilities: [text, tools, json, streaming]
      - name: llama3.2:3b
        context_window: 128000
        capabilities: [text, tools, json, streaming, vision]
      - name: qwen2.5:14b
        context_window: 32000
        capabilities: [text, tools, json, streaming]
      - name: deepseek-r1:1.5b
        context_window: 32000
        capabilities: [text, json, streaming]
      
      # Vision models
      - name: llava:7b
        context_window: 4096
        capabilities: [text, vision, streaming]
      - name: qwen2.5-vl:3b
        context_window: 32000
        capabilities: [text, vision, tools, json, streaming]
      - name: moondream:1.8b
        context_window: 2048
        capabilities: [text, vision, streaming]
    
  # Secondary: Local AirLLM (if installed)
  - name: airllm-local
    type: airllm
    enabled: false  # Enable if pip install airllm
    priority: 2
    models:
      - name: 70b
        default: true
        capabilities: [text, tools, json, streaming]
      - name: 8b
        capabilities: [text, tools, json, streaming]
      - name: 405b
        capabilities: [text, tools, json, streaming]
    
  # Tertiary: OpenAI (if API key available)
  - name: openai-backup
    type: openai
    enabled: false  # Enable by setting OPENAI_API_KEY
    priority: 3
    api_key: "${OPENAI_API_KEY}"  # Loaded from environment
    base_url: null  # Use default OpenAI endpoint
    models:
      - name: gpt-4o-mini
        default: true
        context_window: 128000
        capabilities: [text, vision, tools, json, streaming]
      - name: gpt-4o
        context_window: 128000
        capabilities: [text, vision, tools, json, streaming]
    
  # Quaternary: Anthropic (if API key available)
  - name: anthropic-backup
    type: anthropic
    enabled: false  # Enable by setting ANTHROPIC_API_KEY
    priority: 4
    api_key: "${ANTHROPIC_API_KEY}"
    models:
      - name: claude-3-haiku-20240307
        default: true
        context_window: 200000
        capabilities: [text, vision, streaming]
      - name: claude-3-sonnet-20240229
        context_window: 200000
        capabilities: [text, vision, tools, streaming]

# ── Capability-Based Fallback Chains ────────────────────────────────────────
# When a model doesn't support a required capability (e.g., vision),
# the system falls back through these chains in order.

fallback_chains:
  # Vision-capable models (for image understanding)
  vision:
    - llama3.2:3b          # Fast, good vision
    - qwen2.5-vl:3b        # Excellent vision, small
    - llava:7b             # Classic vision model
    - moondream:1.8b       # Tiny, fast vision
  
  # Tool-calling models (for function calling)
  tools:
    - llama3.1:8b-instruct # Best tool use
    - qwen2.5:7b           # Reliable tools
    - llama3.2:3b          # Small but capable
  
  # General text generation (any model)
  text:
    - qwen3:30b
    - llama3.1:8b-instruct
    - qwen2.5:14b
    - deepseek-r1:1.5b
    - llama3.2:3b

# ── Custom Models ───────────────────────────────────────────────────────────
# Register custom model weights for per-agent assignment.
# Supports GGUF (Ollama), safetensors, and HuggingFace checkpoint dirs.
# Models can also be registered at runtime via the /api/v1/models API.
#
# Roles: general (default inference), reward (PRM scoring),
#        teacher (distillation), judge (output evaluation)
custom_models: []
  # Example entries:
  # - name: my-finetuned-llama
  #   format: gguf
  #   path: /path/to/model.gguf
  #   role: general
  #   context_window: 8192
  #   description: "Fine-tuned Llama for code tasks"
  #
  # - name: reward-model
  #   format: ollama
  #   path: deepseek-r1:1.5b
  #   role: reward
  #   context_window: 32000
  #   description: "Process reward model for scoring outputs"

# ── Agent Model Assignments ─────────────────────────────────────────────────
# Map persona agent IDs to specific models.
# Agents without an assignment use the global default (ollama_model).
agent_model_assignments: {}
  # Example:
  # persona-forge: my-finetuned-llama
  # persona-echo: deepseek-r1:1.5b

# ── Multi-Modal Settings ────────────────────────────────────────────────────
multimodal:
  # Automatically pull models when needed
  auto_pull: true
  
  # Timeout for model pulling (seconds)
  pull_timeout: 300
  
  # Maximum fallback depth (how many models to try before giving up)
  max_fallback_depth: 3
  
  # Prefer smaller models for vision when available (faster)
  prefer_small_vision: true

# Cost tracking (optional, for budget monitoring)
cost_tracking:
  enabled: true
  budget_daily_usd: 10.0  # Alert if daily spend exceeds this
  alert_threshold_percent: 80  # Alert at 80% of budget

# Metrics retention
metrics:
  retention_hours: 168  # Keep 7 days of metrics
  purge_interval_hours: 24