Configuration
Some tools have configurable settings that can be customized per assistant. When a tool supports configuration, a Configure badge appears next to it in the assistant’s Tools tab. Click the badge to open the tool’s configuration page.
Tools ship with sensible defaults, so configuration is optional — you only need to change settings when you want to override the default behavior. A Reset to Defaults button is available to clear any customizations.
search_knowledge
The search_knowledge tool has the following configurable settings:
Search Quality
- Similarity Threshold — Sets the maximum cosine distance (0-2) for search results. Lower values are stricter, filtering out results that aren’t closely related to the query. Default is 1.2 (filters obviously irrelevant results). A value around 0.5 is a good starting point for stricter filtering.
- Source Types to Search — Controls which types of knowledge are included in search results: websites/URLs, facts, documents, conversations, code, and datasets. Uncheck a type to exclude it.
Re-ranking
Re-ranking uses an LLM to score and re-order search results by relevance after the initial vector search. This significantly improves result quality but adds latency and cost.
- Enable Re-ranking — Off by default. When enabled, the tool fetches 20 candidates from vector search, scores each with the LLM, and returns the top 5.
- Re-rank Model — Which model to use for scoring. Default is Claude Haiku 4.5 (fast and cheap). Adds ~200-500ms per search.
- Re-rank Top K — Number of results to return after re-ranking. Default is 5.
- Show Relevance Indicators — When enabled, results include
[High relevance],[Medium relevance], or[Low relevance]labels to help the assistant self-assess confidence. Only works with re-ranking enabled.
Knowledge Boundaries
- Knowledge Boundary Instructions — Rules that constrain the assistant to answer from its knowledge base. These are also configurable from the Knowledge Grounding guardrail. Leave empty to use defaults.
Plugin Tool Configuration
Plugin tools can also define their own configuration schemas. Any configurable plugin tool will show a Configure badge in the same way. See Tool Plugins for how plugin authors define configuration.
Technical Details
Config resolution pipeline — Each tool class can define a config_schema specifying field types, default values, labels, and help text. When an assistant customizes a tool, the overrides are saved per-assistant. At execution time, the system merges the schema defaults with the saved overrides to produce the final configuration — so any field not explicitly customized falls back to its default value.
Runtime loading — All per-assistant tool configurations are loaded once at the start of each agent loop and cached in memory. Before dispatching each tool call, the system injects the correct configuration for that specific tool into the execution context. This means tools always receive their resolved configuration automatically, with no per-call database lookups.