config.yaml

# LLM Directory Configuration

# Request Configuration
timeout: 30.0  # Request timeout in seconds
max_retries: 3  # Maximum number of retries

# Retry Configuration
retry_config:
  max_attempts: 3  # Maximum retry attempts
  base_delay: 1.0  # Base delay between retries (seconds)
  max_delay: 60.0  # Maximum delay between retries (seconds)
  exponential_base: 2.0  # Exponential backoff base
  jitter: true  # Add random jitter to prevent thundering herd

# Rate Limiting Configuration
rate_limit_config:
  requests_per_minute: 600  # Maximum requests per minute
  tokens_per_minute: 150000  # Maximum tokens per minute
  burst_size: 10  # Burst capacity

# Advanced Configuration
fetch_models_on_init: false  # Fetch available models on initialization

# Model Configuration
# Define available models with their capabilities
models:
  # Qwen Models
  Qwen/Qwen3-235B-A22B-Instruct-2507:
    description: Qwen 3 235b A22B Instruct 2507 the updated version of Qwen3 235B
      A22B, with significant improvements in performance. This model is non-thinking.
    input_price_per_1m: 0.13
    max_input_tokens: 256000
    max_output_tokens: 64000
    name: Qwen/Qwen3-235B-A22B-Instruct-2507
    output_price_per_1m: 0.5

  Qwen/Qwen3-235B-A22B-Thinking-2507:
    description: The thinking version of Qwen 3 235b A22B 2507, with enhanced reasoning
      capabilities and step-by-step problem solving.
    input_price_per_1m: 0.3
    max_input_tokens: 256000
    max_output_tokens: 64000
    name: Qwen/Qwen3-235B-A22B-Thinking-2507
    output_price_per_1m: 0.5

  Qwen/Qwen3-Next-80B-A3B-Instruct:
    description: Based on the new Qwen3-Next architecture (hybrid attention, highly
      sparse MoE, training-stability optimizations, and multi-token prediction), the
      Qwen3-Next-80B-A3B-Instruct model delivers extreme efficiency with only 3B active
      parameters per pass. It performs comparably to Qwen3-235B-A22B-Instruct-2507
      and shows clear advantages on ultra-long context tasks (up to 256K tokens).
    input_price_per_1m: 0.15
    max_input_tokens: 256000
    max_output_tokens: 64000
    name: Qwen/Qwen3-Next-80B-A3B-Instruct
    output_price_per_1m: 0.65

  # DeepSeek Models
  deepseek-ai/deepseek-v3.2-exp:
    description: DeepSeek V3.2 Exp, DeepSeek's latest model offering far better performance
      especially on longer contexts than its predecessors. Current flagship model
      by DeepSeek. FP8.
    input_price_per_1m: 0.28
    max_input_tokens: 163800
    max_output_tokens: 40960
    name: deepseek-ai/deepseek-v3.2-exp
    output_price_per_1m: 0.42

  deepseek-ai/deepseek-v3.2-exp-thinking:
    description: DeepSeek V3.2 Exp Thinking, DeepSeek's latest model offering far
      better performance especially on longer contexts than its predecessors. Current
      flagship model by DeepSeek. FP8.
    input_price_per_1m: 0.28
    max_input_tokens: 163800
    max_output_tokens: 40960
    name: deepseek-ai/deepseek-v3.2-exp-thinking
    output_price_per_1m: 0.42

  # Minimax Models
  minimax/minimax-m2.1:
    description: MiniMax M2.1 builds on M2 with enhanced context understanding and
      improved complex tool use. 230B parameter MoE model (10B active) optimized for
      agentic workflows and long-horizon tasks. FP8.
    input_price_per_1m: 0.33
    max_input_tokens: 200000
    max_output_tokens: 50000
    name: minimax/minimax-m2.1
    output_price_per_1m: 1.32

  # Kimi Models
  moonshotai/Kimi-K2-Instruct-0905:
    description: Kimi K2 0905. Kimi-k2 is a Mixture-of-Experts (MoE) foundation model
      with exceptional coding and agent capabilities, featuring 1 trillion total parameters
      and 32 billion activated parameters. In benchmark evaluations covering general
      knowledge reasoning, programming, mathematics, and agent-related tasks, the
      K2 model outperforms other leading open-source models. Quantized at FP8.
    input_price_per_1m: 0.4
    max_input_tokens: 256000
    max_output_tokens: 64000
    name: moonshotai/Kimi-K2-Instruct-0905
    output_price_per_1m: 2.0

  moonshotai/kimi-k2-thinking:
    description: Kimi K2 0905. Kimi-k2 is a Mixture-of-Experts (MoE) foundation model
      with exceptional coding and agent capabilities, featuring 1 trillion total parameters
      and 32 billion activated parameters. In benchmark evaluations covering general
      knowledge reasoning, programming, mathematics, and agent-related tasks, the
      K2 model outperforms other leading open-source models. Quantized at FP8.
    input_price_per_1m: 0.4
    max_input_tokens: 256000
    max_output_tokens: 64000
    name: moonshotai/Kimi-K2-Instruct-0905
    output_price_per_1m: 2.0
    
  moonshotai/kimi-k2.5:
    description: Kimi K2.5 is Moonshot AI's native multimodal model built on Kimi K2
      with ~15T mixed visual and text tokens, delivering strong general reasoning, visual
      coding, and agentic tool-calling. This route uses instant (non-thinking) mode
      for faster responses.
    input_price_per_1m: 0.6
    max_input_tokens: 256000
    max_output_tokens: 64000
    name: moonshotai/kimi-k2.5
    output_price_per_1m: 3.0

  moonshotai/kimi-k2.5:thinking:
    description: Kimi K2.5 with thinking mode enabled. Built on Kimi K2 with ~15T mixed
      visual and text tokens, it excels at general reasoning, visual coding, and agentic
      tool-calling. Produces reasoning traces for complex multi-step workflows.
    input_price_per_1m: 0.6
    max_input_tokens: 256000
    max_output_tokens: 64000
    name: moonshotai/kimi-k2.5:thinking
    output_price_per_1m: 3.0

  # Qwen Coder Models
  qwen/qwen3-coder:
    description: Qwen 3 Coder 480B, a 480 billion total parameter model with 35B active,
      and 160 total experts with 8 active. Performs similar to Claude 4 Sonnet in
      coding benchmarks, but does so at a much lower price.
    input_price_per_1m: 0.13
    max_input_tokens: 262000
    max_output_tokens: 65500
    name: qwen/qwen3-coder
    output_price_per_1m: 0.5

  qwen3-30b-a3b-instruct-2507:
    description: Qwen3-30B-A3B-Instruct-2507 is a 30.5B-parameter mixture-of-experts
      language model from Qwen, with 3.3B active parameters per inference. Significant
      improvements in general capabilities, including instruction following, logical
      reasoning, text comprehension, mathematics, science, coding and tool usage.
    input_price_per_1m: 0.2
    max_input_tokens: 256000
    max_output_tokens: 64000
    name: qwen3-30b-a3b-instruct-2507
    output_price_per_1m: 0.5

  # GLM Models
  z-ai/glm-4.6:
    description: Latest GLM series chat model with strong general performance. Quantized
      at FP8
    input_price_per_1m: 0.6
    max_input_tokens: 200000
    max_output_tokens: 50000
    name: z-ai/glm-4.6
    output_price_per_1m: 1.7

  z-ai/glm-4.6:thinking:
    description: Thinking version of the latest GLM series chat model with strong
      general performance. Quantized at FP8
    input_price_per_1m: 0.6
    max_input_tokens: 200000
    max_output_tokens: 50000
    name: z-ai/glm-4.6:thinking
    output_price_per_1m: 1.7

  zai-org/glm-4.7-flash:
    description: Thinking version of the latest GLM series chat model with strong
      general performance. Quantized at FP8
    input_price_per_1m: 0.6
    max_input_tokens: 200000
    max_output_tokens: 50000
    name: zai-org/glm-4.7-flash
    output_price_per_1m: 1.7

# Chutes Models
  zai-org/GLM-4.7-TEE:
    description: Latest GLM series chat model with strong general performance. Quantized
      at FP8
    input_price_per_1m: 0.6
    max_input_tokens: 200000
    max_output_tokens: 50000
    name: zai-org/GLM-4.7-TEE
    output_price_per_1m: 1.7

  deepseek-ai/DeepSeek-V3.2-Speciale-TEE:
    description: DeepSeek V3.2 Exp Thinking, DeepSeek's latest model offering far
      better performance especially on longer contexts than its predecessors. Current
      flagship model by DeepSeek. FP8.
    input_price_per_1m: 0.28
    max_input_tokens: 163800
    max_output_tokens: 40960
    name: deepseek-ai/DeepSeek-V3.2-Speciale-TEE
    output_price_per_1m: 0.42

  openai/gpt-oss-120b-TEE:
    description: DeepSeek V3.2 Exp Thinking, DeepSeek's latest model offering far
      better performance especially on longer contexts than its predecessors. Current
      flagship model by DeepSeek. FP8.
    input_price_per_1m: 0.28
    max_input_tokens: 131000
    max_output_tokens: 40960
    name: openai/gpt-oss-120b-TEE
    output_price_per_1m: 0.42

  moonshotai/Kimi-K2-Thinking-TEE:
    description: DeepSeek V3.2 Exp Thinking, DeepSeek's latest model offering far
      better performance especially on longer contexts than its predecessors. Current
      flagship model by DeepSeek. FP8.
    input_price_per_1m: 0.28
    max_input_tokens: 131000
    max_output_tokens: 40960
    name: moonshotai/Kimi-K2-Thinking-TEE
    output_price_per_1m: 0.42

  nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16:
    description: Latest GLM series chat model with strong general performance. Quantized
      at FP8
    input_price_per_1m: 0.6
    max_input_tokens: 256000
    max_output_tokens: 64000
    name: nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16
    output_price_per_1m: 0.6

# Langfuse Observability Configuration (optional)
langfuse:
  enabled: true  # Enable/disable Langfuse tracking
  # public_key: pk-lf-...  # Set LANGFUSE_PUBLIC_KEY in .env
  # secret_key: sk-lf-...  # Set LANGFUSE_SECRET_KEY in .env
  # host: http://localhost:3000  # Set LANGFUSE_HOST in .env
  timeout: 5.0  # Request timeout
  enabled_for_litellm: true  # Enable LiteLLM auto-tracking
  enabled_for_langchain: true  # Enable LangChain callback tracking
  enabled_for_custom_client: true  # Enable custom client tracking
  default_tags: []  # Default tags for traces
  prompt_tracking:  # Prompt tracking configuration (optional)
    enabled: true  # Enable/disable prompt tracking in Langfuse (default: false to avoid breaking observability)
    store_in_langfuse: true  # Store prompts in Langfuse (requires enabled: true)
    store_in_postgresql: true  # Store prompts in PostgreSQL (always enabled as fallback)
  cost_tracking:  # Cost tracking configuration (optional)
    enabled: true  # Enable/disable cost aggregation and tracking
    # Model pricing is automatically read from models section above
    # Optional: override pricing here if needed
    # model_pricing:
    #   "qwen/qwen3-coder": 0.0001  # per 1K tokens (example)
  docker:  # Docker Compose service management
    compose_file: langfuse/docker-compose.yml  # Path to docker-compose.yml file
    services: []  # Optional: specific services to manage (empty = all services)

# Workflows Configuration (optional)
# This section configures LangGraph workflow behavior (checkpoints, streaming, etc.)
# For LangGraph workflow LOGGING, see the "Logging Configuration" section below
workflows:
  enabled: true  # Enable/disable workflows support
  checkpoint:  # Checkpoint persistence settings
    enabled: false  # Enable checkpoint persistence
    type: memory  # "memory" (in-memory) or "sqlite" (file-based)
    path: checkpoints.db  # Path for SQLite checkpoints (only used if type is "sqlite")
  human_in_loop:  # Human-in-the-loop settings
    enabled: false  # Enable human-in-the-loop nodes
    timeout: 300.0  # Timeout in seconds for human approval
    approval_required: true  # Require explicit approval (vs auto-approve)
    default_action: reject  # Default action if timeout: "approve" or "reject"
  streaming:  # Streaming settings
    enabled: false  # Enable streaming for graph execution
    stream_mode: values  # "values" (final state), "updates" (state updates), or "debug" (all events)
    include_events: true  # Include event details in stream
  swarm_analysis:  # Swarm Analysis configuration
    enable_tool_calling: true  # Enable tool calling for agents (default: true)
    enable_static_dependency_resolution: true  # Enable static dependency resolution (default: true)
    max_tool_calls_per_agent: 10  # Maximum tool calls per agent (default: 10)

# Role Validation Configuration
role_validation:
  # Safe phrases that contain suspicious words but are legitimate
  # These phrases will always be allowed even if they contain words like "command", "execute", etc.
  safe_phrases:
    - command pattern
    - command pattern validator
    - execution pattern
    - execution context
    - execution flow
    - execution path
    - execution model
    - execution environment
    - system architecture
    - system design
    - system analyst
    - system reviewer
    - architecture pattern
    - design pattern
    - pattern validator
    - pattern analyzer
    - tool execution
    - code execution
    - shell script
    - shell integration
    - root cause
    - root analysis
    - admin panel
    - admin interface
    - admin dashboard

  # Dangerous patterns that should always be rejected (regex patterns)
  # These are context-independent security risks
  dangerous_patterns:
    - (?i)\b(delete|drop|truncate|remove)\s+(all|everything|data|database|table|file|system)\b
    - (?i)\b(execute|exec|eval)\s+(arbitrary|system|shell|command|code)\b
    - (?i)\b(sudo|root)\s+(access|privilege|permission|execute)\b
    - (?i)\b(grant|revoke)\s+(all|admin|root|sudo)\b
    - (?i)\b(alter|update|insert)\s+(system|database|config|security)\b
    - (?i)^(admin|root|sudo|system|shell|command|execute|delete|drop|truncate|remove|alter|update|insert|grant|revoke|eval|exec)\s*$

  # Character normalization map (special chars -> safe replacements)
  char_normalization:
    '!': exclamation
    '#': hash
    $: dollar
    '%': percent
    '&': and
    (: left_paren
    ): right_paren
    '*': star
    +: plus
    /: slash
    <: less
    '=': equals
    '>': greater
    '?': question
    '@': at
    '[': left_bracket
    \: backslash
    ']': right_bracket
    ^: caret
    '{': left_brace
    '|': pipe
    '}': right_brace
    '~': tilde

  # Allowed character pattern (regex)
  # Expanded from original to allow common punctuation
  allowed_char_pattern: ^[a-zA-Z0-9\s\-_&()]+$

  # Maximum role name length
  max_role_name_length: 100

# Prompt Validation Configuration
prompt_validation:
  # Minimum validity score threshold (0.0-1.0)
  # Prompts with overall_score >= this value are considered valid
  min_validity_score: 0.7
  
  # High confidence threshold (0.0-1.0)
  # Prompts with overall_score >= this value are considered high confidence
  high_confidence_threshold: 0.9

# Database Configuration (CodeLumen)
# All database services can be configured via environment variables or this config file
# Environment variables take precedence over config file values
database:
  postgresql:
    host: localhost  # PostgreSQL host (POSTGRES_HOST)
    port: 5432  # PostgreSQL port (POSTGRES_PORT)
    user: postgres  # PostgreSQL user (POSTGRES_USER)
    password: postgres  # PostgreSQL password (POSTGRES_PASSWORD) - CHANGE IN PRODUCTION
    database: codelumen  # Database name (POSTGRES_DB)
    schema: codelumen  # Main schema name (POSTGRES_SCHEMA)
    cache_schema: codelumen_cache  # Cache schema name (POSTGRES_CACHE_SCHEMA)

  redis:
    host: localhost  # Redis host (REDIS_HOST)
    port: 6379  # Redis port (REDIS_PORT)
    password: myredissecret  # Redis password (REDIS_AUTH or REDIS_PASSWORD) - CHANGE IN PRODUCTION
    db: 1  # Redis database number 0-15 (REDIS_DB) - Use DB 1 for CodeLumen (separate from Langfuse)
    key_prefix: 'codelumen:'  # Key prefix for all CodeLumen keys (REDIS_KEY_PREFIX)
    socket_timeout: 5  # Socket timeout in seconds (REDIS_SOCKET_TIMEOUT)
    socket_connect_timeout: 5  # Connection timeout in seconds (REDIS_SOCKET_CONNECT_TIMEOUT)
    health_check_interval: 30  # Health check interval in seconds (REDIS_HEALTH_CHECK_INTERVAL)

  clickhouse:
    host: localhost  # ClickHouse host (CLICKHOUSE_HOST)
    http_port: 8123  # ClickHouse HTTP port (CLICKHOUSE_HTTP_PORT)
    native_port: 9000  # ClickHouse native port (CLICKHOUSE_NATIVE_PORT)
    user: clickhouse  # ClickHouse user (CLICKHOUSE_USER)
    password: clickhouse  # ClickHouse password (CLICKHOUSE_PASSWORD) - CHANGE IN PRODUCTION
    database: codelumen_analytics  # Database name (CLICKHOUSE_DB)
    secure: false  # Use secure (HTTPS) connection (CLICKHOUSE_SECURE)
    verify: true  # Verify SSL certificates (CLICKHOUSE_VERIFY)
    timeout: 30  # Request timeout in seconds (CLICKHOUSE_TIMEOUT)

  minio:
    endpoint: localhost:9090  # MinIO endpoint host:port (MINIO_ENDPOINT)
    access_key: minio  # MinIO access key (MINIO_ACCESS_KEY or MINIO_ROOT_USER)
    secret_key: miniosecret  # MinIO secret key (MINIO_SECRET_KEY or MINIO_ROOT_PASSWORD) - CHANGE IN PRODUCTION
    bucket: codelumen  # Default bucket name (MINIO_BUCKET)
    region: us-east-1  # S3 region (MINIO_REGION)
    secure: false  # Use secure (HTTPS) connection (MINIO_SECURE)
    force_path_style: true  # Force path-style addressing (MINIO_FORCE_PATH_STYLE)

# Logging Configuration
logging:
  enabled: true  # Enable/disable logging
  log_dir: logs  # Directory for log files
  log_level: INFO  # DEBUG, INFO, WARNING, ERROR, CRITICAL
  enable_file_logging: true  # Enable file logging
  enable_console_logging: true  # Enable console logging
  use_json_format: false  # JSON for files (false = readable format)
  date_based_files: true  # app_YYYY-MM-DD.log vs app.log
  rotation:  # Log rotation settings
    max_bytes: 10485760  # 10MB (default 5MB in CodeLumen v2)
    backup_count: 5  # Number of backup files to keep per day
    retention_days: 30  # Keep logs for 30 days (old files will be automatically deleted)
    enable_time_rotation: true  # Enable daily rotation at midnight (works with date_based_files)
  module_loggers:  # Module-specific loggers
    enabled: true  # Enable module-specific loggers
    modules:  # Separate logs for these modules
      - langfuse  # Langfuse integration logs (logs/langfuse/)
      - workflows  # LangGraph workflow execution logs (logs/workflows/)