Active Agent

Appearance

DeepSeek Provider ​

Use DeepSeek through its OpenAI-compatible API. JSON output and tool calling are native; ActiveAgent adapts message roles and other differences from OpenAI's API.

Configuration ​

Basic Setup ​

ruby
class AnalysisAgent < ApplicationAgent
  generate_with :deepseek, model: "deepseek-flash"

  def summarize
    prompt("Summarize this in one sentence: …")
  end
end

Configuration File ​

yaml
# config/active_agent.yml
deepseek:
  service: "DeepSeek"
  api_key: <%= Rails.application.credentials.dig(:deepseek, :api_key) %>
  model: "deepseek-flash"

The service value must be "DeepSeek"; the configuration key can be named differently.

Environment Variables ​

bash
DEEPSEEK_API_KEY=sk-...

DeepSeek has no organization or project scoping; OPENAI_* organization and project variables are ignored.

Models ​

The model defaults to deepseek-flash; DeepSeek's API requires a model.

For the current model list, context windows, and pricing, see DeepSeek's documentation.

DeepSeek's pricing varies by time of day. Thinking tokens are billed even when the answer does not use them.

Thinking Mode ​

DeepSeek enables thinking by default. In one deepseek-flash JSON extraction, the default used 83 output tokens, compared with 7 when thinking was disabled.

ActiveAgent leaves this default unchanged. Opt out per prompt:

ruby
class ColorsAgent < ApplicationAgent
  generate_with :deepseek, model: "deepseek-flash"

  def primary_colors
    prompt(
      "Return the three primary colors as JSON.",
      thinking: { type: "disabled" }
    )
  end
end

DeepSeek ignores temperature, presence_penalty, and frequency_penalty while thinking is enabled. Disable thinking when you need those parameters to apply.

Structured Output ​

DeepSeek supports json_object and json_schema natively:

ruby
class ColorsAgent < ApplicationAgent
  generate_with :deepseek, model: "deepseek-flash"

  # A JSON object, with no schema constraint
  def as_object
    prompt("Return the three primary colors as JSON.", response_format: :json_object)
  end

  # Constrained to a schema
  def as_schema
    prompt("Return the three primary colors.", response_format: :json_schema)
  end
end

See Structured Output for the common format, schema placement, and validation behaviour, which are shared across providers.

Tool Calling ​

Tool calling uses the common tools: format:

ruby
class WeatherAgent < ApplicationAgent
  generate_with :deepseek, model: "deepseek-flash"

  def forecast(city)
    prompt("What is the forecast for #{city}?", tools: [ WeatherTool ])
  end
end

MCP ​

DeepSeek silently ignores mcp_servers; ActiveAgent bridges mcps: as function tools. Add gem "mcp" to use the bridge. See MCP support for configuration and filtering.

Differences from the OpenAI Provider ​

Behaviour
Message rolesDeepSeek accepts system, user, assistant, tool, and latest_reminder. ActiveAgent maps instructions to system because DeepSeek rejects developer with a 422.
temperature and friendsAccepted, but ignored while thinking is on.

See Also ​