Revoxy

Search the corpus

Search across every record kind. Four primitives, and they answer different questions: bm25 (lexical, no model call, fastest), v-rank (semantic, for a described problem), hybrid (both, fused by rank — the default and the most accurate). Returns ranked chunks with their record’s provenance ladder and canonical URL.

One-call grounded context

Search, follow-up reads and the department stance in one call: the best-ranked records whole (within a character budget), the thesis of the function most of them are filed under, and a provenance tally stating the evidence grade of the pack. Use this when the next step is answering from the corpus; use search when the next step is choosing what to read. `omitted` names the ranked records the budget cut.

Get one record

Fetch a record whole by its id — all of its text, its typed refs and its provenance. No search and no model call. This is how you follow a reference or re-read something a search returned only a slice of. A renamed id resolves through the rename table: the record comes back under its live id with `resolved_from` naming the id you asked for — update any stored copy.

Get several records

Fetch up to ten records whole, by id, in one call — the batch follow-up to a search or related result. Ids that do not exist come back in `missing` rather than failing the batch, so nine good ids are never hostage to one stale one. A renamed id resolves through the rename table and comes back under its live id with `resolved_from` set — update any stored copy.

Verify a quote against its record

Check that a quote actually appears in a record before citing it — verbatim up to whitespace and case, against every chunk. Returns the chunk and an anchored URL when found, and an honest miss when not. A match must sit within a single section: a quote spanning a section boundary reports as a miss even though the page reads it contiguously. Presenting corpus text that is not in the corpus is the exact failure this dataset exists to prevent; this is the cheap way to not do it.

List records

Enumerate the corpus, filtered and paginated. Use this to see what exists rather than to find something — search is better at finding. Pass nextCursor back to continue.

Related records

Walk the typed reference graph. "outgoing" is what this record points at (also visible in get_record); "incoming" is what points AT it, which no record knows about itself and is usually the more useful direction. Default both.

Get the taxonomies

The controlled vocabularies in full: functions and sub-functions, industries, core capabilities, ROI hooks, tags. Call this before filtering — the records carry the prose (names, descriptions), and the response’s `filter_values` block carries the exact ids every filter argument accepts, measured from the live index.

Corpus shape at a glance

Counts over the visible corpus — total, and per kind, function, status and provenance — with the index commit. One call answers "what is in here and how much", which otherwise costs a pagination crawl. A facet a record does not claim is not counted as a bucket, so the non-kind breakdowns need not sum to the total. Also the freshness check: `in_sync: false` with a `drift_reason` means the search index and the record store were built from different content commits — search reads the older snapshot

Department theses and their verdicts

Each department’s stated position — what it argues is worth doing and what it argues is not. These are the corpus’s opinions rather than its descriptions, so they are the right thing to quote when asked "what does Revoxy think about X".

Recently reviewed records

Records whose last_reviewed date is on or after `since`, newest first. Note this is REVIEW date, not edit date — the date a human last checked the claim. The corpus records no edit time, so there is no "changed since" to ask for.