Experimental Features

3 min read

This page documents every setting under Experimental Features in the Spaces configuration UI. These features are still being iterated on; behaviour, defaults, and the precise UI surface may change between releases.

Each setting is shown by its UI label, with the underlying configuration field name in italics.

Most experimental features only become visible if the platform has the corresponding feature flag turned on for your tenant. See Feature Flags under Platform / Infrastructure for the current list of Web-Search-specific flags.


Query Review (experimental_features.query_elicitation_config)

Internal field name: query_elicitation_config. Display label in the Spaces UI: Query Review.

A user-facing form that surfaces the proposed search queries before they execute. The user can edit, remove, or add queries, then approve. Useful when transparency about what the assistant is about to search for is important — for example in regulated settings or when users want fine-grained control.

UI Label

Field

Type

Default

Enable Query Review

query_elicitation_config.enable_elicitation

boolean

false

Review Timeout

query_elicitation_config.timeout_seconds

integer (seconds, 1-300)

60

How it works:

  1. The orchestrator AI determines the search queries it wants to run.

  2. Before they execute, a form opens in the chat UI with the proposed queries pre-filled.

  3. The user can edit any query, remove queries, or add new ones, then click Approve (or Decline).

  4. If the user approves, the (possibly modified) queries run and the result flows back to the assistant. If the user declines or the timeout elapses, the search is aborted and the assistant is told why.

Requirements:

  • The Web-Search-specific feature flag must be active for the tenant — see Feature Flags under Platform / Infrastructure for details on enable_elicitation_un_15809 and enable_new_answers_ui_un_14411.

  • When the flag is off, the toggle in the Spaces UI is hidden and the feature has no effect.


Argument Screening (experimental_features.argument_screening_config)

Status: Experimental, next-iteration. Not yet recommended for general production use; behaviour is still being tuned.

An LLM-based pre-screen that runs before the Web Search tool executes. The screen decides whether the proposed arguments (the search query, URLs, etc.) contain sensitive information that should not leave the platform. If they do, the tool call is rejected and the assistant is given a structured rejection message it must surface to the user.

This is a stronger guardrail than the warning paragraph on the root Web Search page — that paragraph relies on user vigilance, while Argument Screening enforces a check programmatically.

UI Label

Field

Type

Default

Enable Argument Screening

argument_screening_config.enabled

boolean

false

Screening Guidelines

argument_screening_config.guidelines

textarea

Built-in default ruleset

System Prompt

argument_screening_config.system_prompt

textarea

Built-in default

User Prompt Template

argument_screening_config.user_prompt_template

textarea

Built-in default ({{ arguments }}, {{ guidelines }})

Rejection Response Template

argument_screening_config.rejection_response_template

textarea

Built-in default ({{ reason }})

Recommended customisation: the Screening Guidelines field is the primary point of customisation for administrators — it states what the screening agent should treat as sensitive. The other prompt fields are advanced and rarely need changes.

Requirements:

  • The Web-Search-specific feature flag must be active for the tenant — see Feature Flags under Platform / Infrastructure for details on enable_web_search_argument_screening_un_18741.

  • When the flag is off, the toggle in the Spaces UI is hidden and the feature has no effect.


Tool Response System Reminder (experimental_features.tool_response_system_reminder)

Status: Experimental, next-iteration. Behaviour is intended to stabilise but the field surface and prompt wording may still change.

When enabled, the Web Search tool attaches a system-reminder message to each successful tool response with citation instructions for the orchestrator. This is in addition to the persistent Tool Format Information For System Prompt that is added to the system prompt at conversation start.

UI Label

Field

Type

Default

Enable Tool Response Reminder

tool_response_system_reminder.enabled

boolean

false

Tool Response System Reminder Prompt

tool_response_system_reminder.system_reminder_prompt

textarea

Same default text as the persistent citation prompt

For the comparison between the two citation fields and when to use which, see Quality and Evaluation.


Why these are grouped here

Each of the three features above is enabled per-Space but is also gated by a platform-level feature flag. The platform decides whether the flag is on for a tenant; space admins decide whether to actually use the feature. That two-layer gating lets the platform keep features in beta without exposing every space admin to in-progress UI surfaces.

Once a feature stabilises, it can move out of Experimental Features and become a first-class section of the Web Search configuration.

Last updated