ares-server 0.2.5

# A.R.E.S Configuration Example
# ==============================
# Copy this file to 'ares.toml' and customize for your deployment.
#
# REQUIRED: Set these environment variables before starting:
#   - JWT_SECRET: A secret key for JWT signing (min 32 characters)
#   - API_KEY: API key for service-to-service authentication
#
# This file demonstrates all available configuration options.
# The server will NOT start without ares.toml present.
#
# Hot Reloading: Changes to ares.toml are automatically detected and applied
# without restarting the server.

# =============================================================================
# Server Configuration
# =============================================================================
[server]
host = "127.0.0.1"                 # Bind address
port = 3000                         # HTTP port
log_level = "info"                  # debug, info, warn, error

# =============================================================================
# Authentication Configuration
# =============================================================================
# For security, secrets are stored in environment variables.
# The config references the ENV VAR NAME, not the actual secret.
[auth]
jwt_secret_env = "JWT_SECRET"       # Name of env var containing JWT secret
jwt_access_expiry = 900             # Access token lifetime (seconds)
jwt_refresh_expiry = 604800         # Refresh token lifetime (seconds)
api_key_env = "API_KEY"             # Name of env var containing API key

# =============================================================================
# Database Configuration
# =============================================================================
[database]
# Local SQLite database (default)
url = "./data/ares.db"

# Turso cloud database (optional - uncomment to enable)
# When both turso_url_env and turso_token_env are set and the env vars exist,
# the server will use Turso instead of local SQLite.
# turso_url_env = "TURSO_URL"
# turso_token_env = "TURSO_AUTH_TOKEN"

# Qdrant vector database for semantic search (optional)
# [database.qdrant]
# url = "http://localhost:6334"
# api_key_env = "QDRANT_API_KEY"    # Optional if Qdrant has no auth

# =============================================================================
# LLM Providers
# =============================================================================
# Define named provider configurations. These are referenced by [models].
# Each provider type requires its Cargo feature to be enabled:
#   - ollama: default
#   - openai: cargo build --features openai
#   - llamacpp: cargo build --features llamacpp

# Ollama - Local inference (no API key required)
[providers.ollama-local]
type = "ollama"
base_url = "http://localhost:11434"
default_model = "ministral-3:3b"

# Multiple Ollama configs with different models
[providers.ollama-fast]
type = "ollama"
base_url = "http://localhost:11434"
default_model = "ministral-3:3b"

[providers.ollama-vision]
type = "ollama"
base_url = "http://localhost:11434"
default_model = "qwen3-vl:2b"

[providers.ollama-code]
type = "ollama"
base_url = "http://localhost:11434"
default_model = "ministral-3:3b"

# Granite model (IBM)
[providers.ollama-granite]
type = "ollama"
base_url = "http://localhost:11434"
default_model = "ministral-3:3b"

# OpenAI API (requires 'openai' feature and OPENAI_API_KEY env var)
# [providers.openai]
# type = "openai"
# api_key_env = "OPENAI_API_KEY"
# api_base = "https://api.openai.com/v1"
# default_model = "gpt-4"

# OpenAI-compatible endpoint (e.g., Azure, local vLLM, Anyscale)
# [providers.azure-openai]
# type = "openai"
# api_key_env = "AZURE_OPENAI_KEY"
# api_base = "https://your-resource.openai.azure.com"
# default_model = "gpt-4-deployment"

# LlamaCpp - Direct GGUF model loading (requires 'llamacpp' feature)
# [providers.llamacpp]
# type = "llamacpp"
# model_path = "./models/granite4-tiny-h.gguf"
# n_ctx = 4096                       # Context window size
# n_threads = 4                       # CPU threads to use
# max_tokens = 512                    # Default max tokens

# =============================================================================
# Model Configurations
# =============================================================================
# Named model configs that reference providers. Agents reference these by name.
# This allows easy model swapping without changing agent configurations.

# Fast model for quick routing decisions
[models.fast]
provider = "ollama-fast"
model = "ministral-3:3b"
temperature = 0.7
max_tokens = 256

# Balanced model for most tasks
[models.balanced]
provider = "ollama-local"
model = "ministral-3:3b"
temperature = 0.7
max_tokens = 512

# Powerful model for complex reasoning
[models.powerful]
provider = "ollama-local"
model = "qwen3-vl:2b"
temperature = 0.5
max_tokens = 1024
# top_p = 0.9
# frequency_penalty = 0.0
# presence_penalty = 0.0

# Granite model (alternative)
[models.granite]
provider = "ollama-granite"
model = "ministral-3:3b"
temperature = 0.7
max_tokens = 1024

# Vision model for multimodal
[models.vision]
provider = "ollama-vision"
model = "qwen3-vl:2b"
temperature = 0.7
max_tokens = 512

# Coding model
[models.coding]
provider = "ollama-code"
model = "ministral-3:3b"
temperature = 0.3
max_tokens = 2048

# =============================================================================
# Tools Configuration
# =============================================================================
# Define available tools and their settings.
# Agents reference tools by name in their 'tools' array.

[tools.calculator]
enabled = true
description = "Performs basic arithmetic operations (+, -, *, /)"
timeout_secs = 10

[tools.web_search]
enabled = true
description = "Search the web using DuckDuckGo (no API key required)"
timeout_secs = 30

# Example: Database query tool (not implemented by default)
# [tools.database_query]
# enabled = false
# description = "Execute read-only database queries"
# timeout_secs = 60

# =============================================================================
# Agent Configurations
# =============================================================================
# Each agent has a model, optional tools, and a system prompt.
# The system_prompt can be customized to change agent behavior.

[agents.router]
model = "fast"
tools = []
max_tool_iterations = 1
parallel_tools = false
system_prompt = """
You are a routing agent that classifies user queries.

Available agents:
- product: Product information, catalog, inventory
- invoice: Billing, payments, invoices
- sales: Sales metrics, revenue, customers
- finance: Financial analysis, budgets, expenses
- hr: HR policies, employees, benefits
- orchestrator: Complex multi-domain queries

Respond with ONLY the agent name (one word, lowercase).
"""

[agents.orchestrator]
model = "powerful"
tools = ["calculator", "web_search"]
max_tool_iterations = 10
parallel_tools = false
system_prompt = """
You are an orchestrator agent for complex queries.

Capabilities:
- Break down complex requests
- Delegate to specialized agents
- Perform web searches
- Execute calculations
- Synthesize multiple results

Provide comprehensive, well-structured answers.
"""

[agents.product]
model = "balanced"
tools = []
max_tool_iterations = 5
system_prompt = """
You are a Product Agent for product-related queries.

Capabilities:
- Product catalog and search
- Specifications and details
- Inventory and availability
- Comparisons and recommendations
- Pricing information
"""

[agents.invoice]
model = "balanced"
tools = ["calculator"]
max_tool_iterations = 5
system_prompt = """
You are an Invoice Agent for billing queries.

Capabilities:
- Invoice lookup and status
- Payment processing
- Billing history
- Payment terms
- Discrepancy resolution
"""

[agents.sales]
model = "balanced"
tools = ["calculator"]
max_tool_iterations = 5
system_prompt = """
You are a Sales Agent for sales analytics.

Capabilities:
- Sales performance metrics
- Revenue analysis
- Customer acquisition
- Pipeline insights
- Commission calculations
"""

[agents.finance]
model = "balanced"
tools = ["calculator"]
max_tool_iterations = 5
system_prompt = """
You are a Finance Agent for financial analysis.

Capabilities:
- Financial statements
- Budget tracking
- Expense management
- Cash flow analysis
- Cost accounting
"""

[agents.hr]
model = "balanced"
tools = []
max_tool_iterations = 5
system_prompt = """
You are an HR Agent for human resources queries.

Capabilities:
- Employee information
- Company policies
- Benefits and compensation
- Leave management
- Recruitment
"""

# =============================================================================
# Workflow Configurations
# =============================================================================
# Define how requests flow through agents.

[workflows.default]
entry_agent = "router"              # First agent to handle requests
fallback_agent = "orchestrator"     # Fallback if routing fails
max_depth = 3                       # Max recursive depth
max_iterations = 5                  # Max iterations per workflow

[workflows.research]
entry_agent = "orchestrator"
max_depth = 3
max_iterations = 10
parallel_subagents = true           # Execute subagents in parallel

# =============================================================================
# RAG (Retrieval Augmented Generation) Configuration
# =============================================================================
[rag]
embedding_model = "BAAI/bge-small-en-v1.5"
chunk_size = 1000                   # Characters per chunk
chunk_overlap = 200                 # Overlap between chunks

# =============================================================================
# Dynamic Configuration Paths (TOON Files)
# =============================================================================
# ARES uses a hybrid TOML + TOON configuration approach:
# - TOML (this file): Static infrastructure config (server, auth, database, providers)
# - TOON (config/*.toon): Dynamic behavioral config (agents, workflows, models, tools, MCPs)
#
# Benefits of TOON format:
# - 30-60% token savings when sending to LLMs
# - Hot-reloadable without server restart
# - One file per entity for easy management
# - LLM-friendly format optimized for AI consumption
#
# See docs/TOON.md for more information on the TOON format.

[config]
agents_dir = "config/agents"       # Agent definitions (*.toon)
workflows_dir = "config/workflows" # Workflow definitions (*.toon)
models_dir = "config/models"       # Model configurations (*.toon)
tools_dir = "config/tools"         # Tool configurations (*.toon)
mcps_dir = "config/mcps"           # MCP server configurations (*.toon)
hot_reload = true                  # Watch for changes and reload
watch_interval_ms = 1000           # How often to check for changes