ToolSet

Defined in: src/lib/tools/serverTools.ts:812

A tool set defines a group of tools that work together. When any “anchor” tool in the set is matched semantically (with a score at or above anchorMinSimilarity), the set is activated and all of its members are pulled into the selection.

Two activation strategies consume this interface:

expandToolSetsAdditive (used by useChatStorage and createServerToolsFilter) keeps all original matches and adds the set’s members on top — non-set tools are never dropped.
applyToolSets is exclusive: it keeps only set members plus non-set tools that scored above independentThreshold.

Pick expandToolSetsAdditive when you want recall over precision (typical), and applyToolSets when you specifically want non-set matches stripped on activation.

Properties

anchorMinSimilarity?

optional anchorMinSimilarity: number

Defined in: src/lib/tools/serverTools.ts:828

Minimum similarity an anchor must reach to activate the set. Prevents false activation on prompts where the anchor barely passes the global minSimilarity threshold. Default: 0.60

anchors

anchors: string[]

Defined in: src/lib/tools/serverTools.ts:822

Tools that trigger the set when selected. If any anchor appears in the semantic match results with a score at or above anchorMinSimilarity, all members are pulled in.

members

members: string[]

Defined in: src/lib/tools/serverTools.ts:816

All tool names in the set

name

name: string

Defined in: src/lib/tools/serverTools.ts:814

Human-readable name for logging/debugging

systemPrompt?

optional systemPrompt: string

Defined in: src/lib/tools/serverTools.ts:839

System-prompt fragment to APPEND to the base prompt when this set activates. Additive, never a replacement — it composes with the host’s persona / memory. Gated on genuine activation (anchor score ≥ anchorMinSimilarity, or a forced set) via activatedToolSetNames → toolSetSystemPrompts, NOT on mere anchor presence — a borderline anchor kept by recall-over-precision must not drag this persona in. Write it to be self-limiting too (condition its behavior on the user actually wanting what the set does), so a borderline activation doesn’t bias the turn.