Skip to content

seaticket_config.yaml

Besides .env for general settings, the seaticket_config.yaml is used to specify the detailed configurations of the all components in SeaTicket docker.

File structure of seaticket_config.yaml

seaticket_config.yaml consists of global and component-specific sections. Use seaticket for the web component (without -web) and seaticket-* for the other components. Internal setting names such as SEAQA_* remain unchanged, except for SEATICKET_CONFIG_NAME, SEATICKET_LOG_LEVEL, and SEATICKET_LOG_TO_STDOUT.

For an initial deployment, only global.LLM_MODELS and global.EMBEDDING_MODEL are required. All other settings and component-specific sections are optional and can be omitted to use their defaults.

global:
  LLM_MODELS:
    # ... at least one valid LLM configuration (required)
  EMBEDDING_MODEL:
    # ... embedding model configuration (required)
  # ... other global settings (optional)
seaticket-ai: # optional
  # ... special configurations for seaticket-ai
seaticket-events: # optional
  # ... special configurations for seaticket-events
seaticket-indexer: # optional
  # ... special configurations for seaticket-indexer
seaticket-notification: # optional
  # ... special configurations for seaticket-notification
seaticket: # optional
  # ... special configurations for seaticket
seaticket-io: # optional
  # ... special configurations for seaticket-io

Global settings

The global section contains settings shared by all SeaTicket components. Configure the LLMs and embedding model below before starting SeaTicket; the remaining global settings are optional.

LLMs configurations

To ensure the SeaTicket service functions normally, please ensure that seaticket_config.yaml contains at least one valid LLM

For LLMs configurations, you should configure a list option LLM_MODELS in seaticket_config.yaml:

global:
  LLM_MODELS:
    # <model 1 configuration>
    - type: 'your_llm_provider' # required, please refer https://docs.litellm.ai/docs/providers for the details
      model: 'your_model_id' # required
      url: 'llm_endpoint' # required for some provider (or type)
      key: 'llm_secret_key' # required for non-self deploy LLM server (including OpenAI proxy)
      label: 'display_name' # optional, display name on SeaTicket chat page, default to option `model` if not set
      hidden: false # optional, hidden on SeaTicket chat page but still can be used with API calling and other components, default to `false`
      disable: false # optional, disable this LLM, default to `false`
      tier: # optional, accepts one of three values — `low`, `medium`, or `high` — and is used to specify the model's inference strength, allowing for flexible invocation across different tasks.
      price: # required only need to specify this when you need to track user token usage.
        input_tokens: 1 # dollars / 1 M tokens (non-cached)
        cached_tokens: 0.1 # dollars / 1 M tokens (cached)
        output_tokens: 10 # dollars / 1 M tokens

    # <model 2 configuration>
    # - type: ...

    # <model 3 configuration>
    # - type : ...

Futher information about tier of a model

  • When multiple LLMs have the same tier, the system will use only the LLM that explicitly specifies that tier, with priority given to the first entry in the LLM_MODELS list.
  • If the tier specified for a service in AI_UTILS_TIER is not defined in LLM_MODELS, then the default model in LLM_MODELS or the first valid model in the list will be used.

LLMs reference configuration

global:
# ...
LLM_MODELS:
  - type: openai
    model: <your model-id>  # e.g., gpt-4o-mini
    key: <your openai LLM access key>
global:
# ...
LLM_MODELS:
  - type: gemini
    model: <your model-id>  # e.g., gemini-2.5-flash
    key: <your gemini LLM access key>
global:
# ...
LLM_MODELS:
  - type: dashscope
    model: <your model-id>  # e.g., qwen3-max
    key: <your qwen LLM access key>
global:
# ...
LLM_MODELS:
  - type: deepseek
    model: <model_id> # e.g. deepseek-chat
    key: <your LLM access key>
global:
# ...
LLM_MODELS:
  - type: azure
    model: <your deployment name>
    url: # your deployment url, leave blank to use default endpoint
    key: <your LLM access key>
global:
# ...
LLM_MODELS:
  - type: ollama
    model: <your model-id>
    url: <your LLM endpoint>
    key: <your LLM access key>
global:
# ...
LLM_MODELS:
  - type: huggingface
    model: <model provider>/<model-id>
    url: <your huggingface API endpoint>
    key: <your huggingface API key>

SeaTicket AI utilizes LiteLLM to interact with LLM services. For a complete list of supported LLM providers, please refer to this documentation. Then fill the following fields in your seaticket_config.yaml:

For example, if you are using a LLM service with OpenAI-compatible endpoints, you should set type to other, and set other LLM configuration items accurately.

Embedding model

To ensure the SeaTicket service functions normally, please ensure that seaticket_config.yaml has specified embedding model

The configuration of embedding models is similar to that of an element in the LLM_MODELS list:

global:
  EMBEDDING_MODEL:
    type: ...
    model: embedding_model_id
    url: embedding_model_endpoint # required for some provider (or type)
    key: embedding_model_secret_key # required for non-self deployment LLM server (including OpenAI-proxy)
    price: 0.02 # dollars / 1 M input tokens, required only need to specify this when you need to track user token usage.

Internal server URLs (Optional)

For the standard single-host Docker deployment, no internal endpoint configuration is required. Components use these built-in defaults:

Setting Default Used by
SEAQA_WEB_INNER_SERVER_URL http://seaticket/ AI and Notification
SEAQA_AI_INNER_SERVER_URL http://seaticket-ai:8887 Web and Indexer
SEAQA_INDEXER_INNER_SERVER_URL http://seaticket-indexer:8888 Web, AI, and Events

The web image exposes Nginx on port 80. Gunicorn listens only on 127.0.0.1:8000 inside that container, so do not use http://seaticket:8000/ for communication from another container.

For a custom or multi-host deployment, override the relevant URLs in seaticket_config.yaml using the existing SEAQA_* names:

global:
  # ... other configurations
  SEAQA_WEB_INNER_SERVER_URL: http://<your seaticket server host>/
  SEAQA_AI_INNER_SERVER_URL: http://<your seaticket-ai server host>:8887
  SEAQA_INDEXER_INNER_SERVER_URL: http://<your seaticket-indexer server host>:8888

SEAQA_EVENTS_INNER_SERVER_URL is retained in the web configuration with a default of http://seaticket-events:6001, but the current Events component runs background workers and does not expose an HTTP listener on port 6001. No Events HTTP endpoint needs to be configured for the standard deployment.

Thought process (Optional)

The thought process is used to display detailed debug information about LLM calls, including system prompts, context, tool call history, and cost statistics. It will be displayed in the project's Chat and Agent modules.

global:
  THOUGHT_PROCESS: # Configure which content to record in the thought process.
    enabled: false # set to `true` to record in the thought process (global switcher)
    tasks:
      enabled: true
      system_prompt: false
      user_input: true
      user_input_raw: false
    context: false
    tool_details: false

Thirdparty integrations (Optional)

Please refer here for the details.

seaticket-ai settings (Optional)

The following are the default values, normally, you don't need to make any additional settings for this part

seaqa-ai:
  MAX_STEPS: 20 # Maximum Chat execution steps per request
  AGENT_MAX_STEPS: 20 # Default maximum execution steps per Agent phase
  MAX_TOOL_CALLS_PER_RESPONSE: 3 # Maximum tool calls executed from one Chat model response
  GET_RECORDS_MAX_RECORDS: 10 # Shared successful record-reading budget per Chat request or complete Agent task
  SEARCH_TOOL_CONTENT_MAX_SIZE: 20000 # Maximum content length per record, in characters (not tokens)
  SEARCH_TOOL_ISSUES_MAX_COMMENTS: 50 # Maximum issue comments included when reading a record
  COMPLETION_MAX_RETRIES: 2 # Maximum number of chating completion retries with LLM
  COMPLETION_RETRY_INTERVAL: 0.0 # Retry interval when AI chating completion fails
  TOOL_CALL_MAX_RETRIES: 2 # Maximum number of executing tool retries with LLM
  TOOL_CALL_RETRY_INTERVAL: 1.0 # Retry interval when tool executes fails
  AI_UTILS_TIER: # The tier level required to configure AI functionality (must be one of `low`, `medium`, and `high`) will be matched in `LLM_MODELS`.
    generate_ticket: low
    generate_summary: low
    rerank_relevance: medium

Execution and record-reading limits

Setting Default Scope
MAX_STEPS 20 Maximum execution steps for a Chat request. Existing explicit configuration overrides this default.
AGENT_MAX_STEPS 20 Default maximum execution steps for each Agent phase, not for the entire task. A phase can override this limit.
MAX_TOOL_CALLS_PER_RESPONSE 3 Maximum tool calls executed from a single Chat model response. This is separate from the execution-step limit.
GET_RECORDS_MAX_RECORDS 10 Maximum distinct records successfully read and provided per Chat request or complete Agent task, shared across all Agent phases.
SEARCH_TOOL_CONTENT_MAX_SIZE 20000 Content budget for a single record read by get_records, measured in characters, not tokens.
SEARCH_TOOL_ISSUES_MAX_COMMENTS 50 Maximum issue comments included when reading a record. The content budget still applies.

About GET_RECORDS_MAX_RECORDS

Search tools return metadata, summaries, and available snippets rather than full record content. Chat and Agent use get_records(targets=[...]) to selectively read records. Repeated reads of the same record do not consume additional record budget; invalid targets and failed reads do not consume it either. Once the record budget is exhausted, additional records cannot be read, even if execution steps remain.

About MAX_STEPS

MAX_STEPS, AGENT_MAX_STEPS, and MAX_TOOL_CALLS_PER_RESPONSE must be positive integers. These are upper limits: Chat and Agent can finish earlier when enough evidence is available. Increasing execution limits can increase latency and token usage; it does not increase the shared record-reading budget.

Model selection

  • For the chat and agent, the models typically used include:
    • Main resonance chain:
      • Chat: The model selected for the webpage (LLM_MODELS.hidden = false)
      • Agent: The defualt model in LLM_MODELS
    • An embedding model for vector search
    • A model specified by AI_UTILS_TIER.rerank_relevance for reranking search results
  • For other services (e.g., summary generation), will use the model specified in AI_UTILS_TIER

Agent configurations

seaticket-ai:
  AGENT:
    # --- Max records counter ---
    max_tickets_per_run: 100 # Maximum tickets analyzed per agent run
    max_github_issues_per_run: 100 # Maximum GitHub issues analyzed per agent run

    # --- Agent manager timing (seconds) ---
    enqueue_pending_events_check_intetval: 5 # Interval to check and flush pending events
    enqueue_debounce_interval: 600 # Minimum seconds between two enqueues for the same record
    enqueue_ticket_initial_debounce: 300 # Initial debounce seconds after first ticket event
    timed_scan_interval: 3600 # Timed scanner execution interval
    timed_scan_project_cooldown: 86400 # Cooldown before re-scanning the same project

seaticket-events settings (Optional)

seaticket-events:
  AI_STATS: # optional
    enabled: true # Enable AI tokens usage statistics (default is `true`)

seaticket notification settings (Optional)

The Docker examples set this value through .env.

seaticket:
  ENABLE_NOTIFICATION_SERVER: false