Skip to main content
The payload view: takes the top-ranked hits and assembles a markdown block — per hit, ## title (source · date · rel 0.84), then the record’s path (a source_native_id key into brain.bronze.raw_records, so an agent can drill into the full record via duckdb), then the snippet — filling top-down until the character budget is reached. Designed to be pasted into a prompt or consumed directly by an agent. search is for scanning and picking; context is for feeding a model. Same facets and modes apply to both, but context is tuned for injection: it refuses low-information prompts (fewer than 2 content terms), defaults to --mode hybrid, excludes the agent’s own session transcripts (--include-sessions to override), deduplicates identical snippets, and applies a relevance floor (--min-relevance, default 0.55) — injecting nothing beats injecting noise, and an empty result is a valid answer. --json emits {query, budget, context} on one line — what every host plugin’s hook parses (see For Agents).