Agentic AI Settings

Fine-tune token budgets, model behaviour, guardrails, and AI usage limits for the agentic pipeline and workflow execution engine.

AI Usage Limits

These settings cap how many resources the AI engine may consume per session or per day, preventing runaway costs from long-running pipelines.

Option	Description
Max tokens per request	Hard limit on the number of output tokens generated per individual AI call. Requests that would exceed this are truncated.
Max requests per session	Total number of AI calls allowed within a single browser session. Prevents unintended loops.
Daily token budget	Aggregate token ceiling across all AI calls per calendar day. Resets at midnight UTC.

Agent Behaviour

Option	Values	Description
Retry on failure	0 – 5	How many times to automatically retry a failed AI call before surfacing an error to the user.
Temperature	0.0 – 1.0	Controls output creativity. Lower values produce more deterministic, factual text. Higher values introduce more variation and creativity.
Streaming	On / Off	When enabled, AI output streams token-by-token so the user sees text appearing in real time rather than waiting for the full response.
Enable logging	On / Off	Logs every AI request and response to the server log file. Useful for debugging but increases disk usage.

Temperature guide

For factual tasks (data summaries, code analysis) keep temperature at 0.2 – 0.4. For creative writing (blog posts, stories) raise it to 0.7 – 0.9.

Guardrails

Guardrails add a layer of safety between raw user input / AI output and the rest of the application.

Guardrail	Description
Input sanitisation	Strips potentially harmful patterns (prompt injection attempts, HTML script tags) from user input before it is sent to the AI model.
Output validation	Checks AI-generated content against a blocklist of prohibited terms. Blocked responses are replaced with a safe fallback message.
PII detection	Scans AI output for patterns that look like personal data (email addresses, phone numbers, credit card numbers) and redacts them before display.

Note

Guardrails add a small latency overhead per request. Disable PII detection only if your prompts and outputs are guaranteed to contain no personal data.

AI Engine Selection

Choose which AI model is used for agentic workflow steps. This is separate from the Terminal Writer Engine — it applies to workflow node execution, task analysis, and assistant panels.

Engine	Description
Hosted API (default)	Calls the Anthropic Claude API (`claude-sonnet-4-6`) via the server. Requires `ANTHROPIC_API_KEY` in the server environment.
Local CLI	Uses the locally installed `claude` CLI binary. Suitable when API access is restricted.
Local LLM	Uses an on-device BERT model. Supports classification tasks only — generation is not available.

Embedding Models (Ollama)

The Embedding Engine section of this tab lets you download and select local Ollama models used for Semantic Search and knowledge ranking — separate from the text-generation engine above.

FAQ

Does the daily token budget reset at my local midnight or UTC?

It resets at midnight UTC, regardless of the user's timezone setting. Plan heavy batch operations accordingly.

What happens when the daily token budget is exhausted?

All AI calls return an error message stating the budget has been reached. The budget resets automatically at midnight UTC — no manual reset is required.

Does enabling logging store the content of AI responses?

Yes — full request and response payloads are written to the server log. Disable logging if your prompts contain sensitive business data you do not want persisted on disk.