Skip to main content

Researcher Agent Config (Optional)

Use RESEARCHER_AGENT_CONFIG to define the LLM stacks that power the dedicated research workflow. The JSON payload should describe each specialized role plus the shared credentials it needs. Populate this variable only when you want II-Agent to spin up the researcher/report builder chain.

RESEARCHER_AGENT_CONFIG='{
  "final_report_builder": {
    "model": "gemini-2.5-pro",
    "application_model_name": "gemini-2.5-pro",
    "api_key": "<GEMINI_API_KEY>",
    "base_url": null,
    "max_retries": 3,
    "max_message_chars": 30000,
    "temperature": 0.0,
    "vertex_region": null,
    "vertex_project_id": null,
    "api_type": "gemini",
    "thinking_tokens": 0,
    "azure_endpoint": null,
    "azure_api_version": null,
    "cot_model": false
  },
  "report_builder": {
    "model": "gemini-2.5-flash",
    "application_model_name": "gemini-2.5-flash",
    "api_key": "<GEMINI_API_KEY>",
    "base_url": null,
    "max_retries": 3,
    "max_message_chars": 30000,
    "temperature": 0.0,
    "vertex_region": null,
    "vertex_project_id": null,
    "api_type": "gemini",
    "thinking_tokens": 0,
    "azure_endpoint": null,
    "azure_api_version": null,
    "cot_model": false
  },
  "researcher": {
    "model": "deepseek-reasoner",
    "application_model_name": "r1",
    "api_key": "<DEEPSEEK_API_KEY>",
    "base_url": "https://api.deepseek.com/beta",
    "api_type": "openai"
  }
}'

Implementation notes​

  • Builders need full configs. Both report_builder roles must include the full LLM configuration block (model, application model name, API key, retry/size limits, and toggles) so they can render structured drafts without reusing the base agent credentials. Customize temperature, max_message_chars, or any other fields to match your latency/cost requirements.
  • Researcher must use the completion API. The researcher role runs long-form reasoning, so wire it to an OpenAI-compatible completion endpoint (DeepSeek, OpenAI, Anthropic-compatible shim, etc.) and provide api_type + base_url as needed.
  • Scope secrets carefully. Store the JSON in .stack.env (never in version control) and restart the backend service after edits so the new config loads into the process environment.