kibi-plugin-search-hints
v1.0.1
Published
Offline meaning-based hints for weak Kibi search answers
Readme
Kibi search hints plugin
This package provides the kibi.search-hints.v1 capability. When kb_search gives a weak or abstained lexical answer, Kibi asks it which current entities the question is probably about, and shows up to ten of their requirements under queryAnalysis.hints and in one sentence of the text answer. The agent then re-asks in the KB's own words. Hints never change search results, the governing answer, abstention, kb_check or proofs.
Activate it in the project's package.json and install it as a dev dependency:
{
"kibi": {
"plugins": [
{
"package": "kibi-plugin-search-hints",
"capabilities": { "kibi.search-hints.v1": { "mode": "replace" } }
}
]
}
}Use "mode": "shadow" to compute hints for comparison without showing them. kb_search with hints: false skips them for one call.
The plugin embeds text with the bundled potion-base-8M static embedding (256 dimensions, int8, about 7.7 MB): a text's vector is the normalised mean of its tokens' table rows. It is pure TypeScript with no native code and no network access, and the same text gives the same vector on every platform. Entity vectors are cached under the directory Kibi passes in, keyed by the model and each entity's content hash, so only changed entities are re-embedded. The model is English-only.
See catalog.json, integrity.json, SBOM.spdx.json and THIRD_PARTY_NOTICES.md for the shipped component record.
