Skip to content
KallamoDocsBack to the site

Smart search

Every reply starts with a search for relevant context. By default that search uses your message as it is: a plain similarity search against its text. Smart search lets the AI read the conversation first and write its own searches, follow leads, and stop when it has enough.

Turn it on in the profile, under Smart search › Search before replying.

Smart search on Smart search off
Questions about the canon: “who was at the festival last year?” Short back-and-forth: “continue”, “ok”, “make it shorter”.
Long stories where the answer is spread across many places. Editing a passage that holds everything the AI needs.
A game master helper, a continuity checker, a lore keeper. A quick brainstorm with no canon to respect.

Smart search costs extra requests on the Retrieval planner model (AI roles), so it pays to give that role a fast, inexpensive model. If the role is off, the profile falls back to the plain search.

Search rounds (1 to 5, default 3). One round is a search and a look at what came back. 1 is cheapest, 3 is balanced, more follows leads further and costs more.

Search instructions tell the search step what to look for. The reply itself still follows Behavior and rules. For example:

Find the passages where the selected characters and places appear, so the edit keeps every fact consistent.

In each round the planner can:

  • search the profile’s knowledge base, the workspace’s files and memory, and the archive of older chat;
  • read a whole file when a passage is not enough;
  • look up a Worldbuild entity, and follow its relationships;
  • recall an archived part of the conversation.
  • It never starts from nothing. The plain search runs first, for free, and its results are already in the context before the planner is asked anything. Smart search can only add to it.
  • It skips messages that need no research. A short follow-up with no name and no question uses the plain search. Plan every message, in Settings › AI roles, turns this off.
  • It never repeats a search it already ran, in this reply or in recent ones.
  • It stops when the turn limit or the budget is reached, and keeps everything it found.

The planner uses the provider’s own tool calling when the connection supports it (OpenAI, OpenRouter, compatible local servers, Anthropic, Claude on Bedrock, Google AI, Vertex AI), and a text protocol everywhere else. It switches on its own if native tools fail. Native tool calling, in Settings › AI roles, turns native tools off.