Settings
Keys stay in this browser (`localStorage`). They are sent only to KAI API routes for your requests and are never written to the database.
Required for indexing and chat unless a server `GOOGLE_API_KEY` fallback is configured.
Models
Retrieval
Dense + BM25, fused with weighted RRF. Best general default.
Raise sparse for technical corpora full of exact identifiers; raise dense for prose where wording varies.
A second pass scores each (question, passage) pair, which removes candidates that fused highly on vocabulary overlap alone.
Answer quality
Each claim is re-checked against the passages it cites. Unsupported and miscited claims are flagged in the answer.
Below the threshold KAI reports what it found and what is missing instead of generating an answer. Set to 0 to always answer.
Indexing
Recursive split that respects section headings and keeps the heading breadcrumb.
Stops repeated boilerplate across documents from crowding out the passages that answer the question.