RAG

/rag is retrieval-augmented question answering built entirely out of ordinary agents: no separate RAG service, no bespoke answer UI. Ask a question and the agent files its answer as child bullets under the node you asked from, each claim bold and backed by a [[source page]] link you can click straight to the evidence.

Three strategies, three agents

Boot ensures three editable definition pages under agents/, each wrapping a different retrieval strategy over the same knowledge base. Pick whichever fits the question, or bind your own agent to one of these as a starting point.

Agent	Strategy	Underlying tool
/rag	Hybrid-RRF baseline: hybrid search results, then reads the top sources	kb.search
/rag-pageindex	Agentic directory expansion: walks the directory tree one level at a time, PageIndex-style, deciding where to descend	kb.expand
/rag-direct	Direct corpus integration: a bounded, title-ordered dump of the whole readable corpus, no ranking involved	kb.corpus

/rag, hybrid-RRF baseline

Runs kb.search (the same reciprocal-rank-fused lexical-plus-vector search behind ⌘K), then reads the top-ranked source pages before drafting an answer. This is the default choice for most questions: fast, cheap, and grounded in whatever ranks highest.

/rag-pageindex, agentic directory expansion

Calls kb.expand {directory?} one level at a time, starting from the workspace root, and decides where to descend based on what each level returns. Better than the baseline when the answer depends on structure, “everything under research/”, rather than on what a ranked search happens to surface.

/rag-direct, direct corpus integration

Calls kb.corpus {maxChars?}, a bounded, title-ordered concatenation of every readable node grouped under [[source page]] headings. It bypasses embedding and rank retrieval entirely, so it’s the right choice for small corpora or exhaustive questions where you don’t trust a ranking to have surfaced everything relevant.

Asking a question

Each /rag agent’s argsSchema accepts one required question argument, a full natural-language question, not a keyword query. Invoke it like any slash agent:

/rag what did we conclude about warm-start clustering in the segmentation project?

The agent is prompted to produce one bold claim per output child bullet, each with a [[source page]] link back to where that claim came from:

**Warm-start clustering cut convergence time by roughly 40% on the crisp-segmentation baseline.** [[research/crisp-segmentation-profile]]
**The effect held only when the warm-start centroids came from the prior week's run.** [[research/warmup-hypothesis]]

Answers file into the invoking page exactly like any other agent run: violet output bullets grow under the node you invoked from, and the run’s introspection chips let you inspect exactly what got retrieved.

Asking from the palette

The ⌘K palette appends an “Ask /rag” row below the ordinary hybrid search results. Pressing ⇧↵ on it files /rag <your query> on the page you have open and starts the same child-node run there, so you never have to leave what you’re reading to get a sourced answer.

Note

This is the fast escalation path from search to retrieval: type a query, see the ranked hits, and if none of them directly answer the question, ⇧↵ turns the query into a full RAG run without retyping it. See search for the palette and its ranked-hit behavior.

Related

SearchHybrid lexical-plus-vector search, reciprocal-rank fusion, and the ⌘K palette these agents build on.

Knowledge ingestionHow the compiled wiki these agents retrieve over gets built and kept current.

Agent definitionsThe definition shape, argsSchema, and slash invocation mechanics /rag uses.

Agents overviewThe run tree and output bullets that render a RAG answer under the invoking node.