Filter and collect, not rank
A ranked search asks “what are the best matches?”. A sweep asks “which units match?” and returns all of them — so completeness is a property of the query rather than a hope about the value of k.
Exhaustive Find
Differentiator“All” should mean all.
Ranked search returns an answer that looks complete and is quietly partial. The exhaustive tier sweeps the whole access-scoped index and reports exactly what it read, what it is still waiting on, and what it could not read.
01What it is
Ask for every mention of a party across a data room and ranked search will return ten good passages. The answer looks complete and is quietly partial — the worst failure mode in legal work. Exhaustive Find is a different operation: a complete sweep of the access-scoped index, with a manifest of what it read.
02In detail
A coverage manifest, always
Every run records each document as scanned, awaiting index, or failed. A document still being indexed is waited on and resumed — never silently dropped, which would recreate the exact failure this exists to prevent.
It follows the reference, not just the name
Defined terms and aliases are resolved at ingestion, so “the Purchaser” reaches the same entity as the party's name — deterministically, with each mention tagged by how it was resolved.
Bounded and scoped
Sweeps are container-scoped and access-filtered before anything else, run against your firm's budget, and refuse with a clear message above a configured size rather than running away with your spend.
03In practice
Completeness
312 of 314 read. Here are the other two.
A sweep that quietly dropped a document would be indistinguishable from a complete one. So it never does — it reports.
Exhaustive find · coverage manifestrun 4f21c8
Project Atlas — data room (287 files)12,904 passagesScanned
Disclosure letter (v7).docx96 passagesScanned
Supply agreement — Meridian.pdf412 passagesScanned
Side letter — Meridian.emlindexingAwaiting
Board pack 2024 (scan).pdfOCR failedFailed
312 of 314 documents read41 mentions foundComplete
- Scanned
- 312
- scanned
- Awaiting index
- 1
- awaiting index
- Failed
- 1
- failed
A document still being indexed is waited on, never skipped — the run parks and resumes when ingestion finishes.
A sweep that quietly dropped a document would be indistinguishable from a complete one. So it never does: the exceptions are the output.
ScannedAwaitingFailed
04Why it matters
How it fits the matter
It is the difference between telling a client you looked and being able to show them exactly what you read — and what you could not.
05Questions
Straight answers
What people ask before they buy this.
- How is this different from just raising the number of results?
- Raising k still trusts a similarity ranking to have put everything relevant near the top. A sweep does not rank at all — it filters the whole access-scoped index and collects every unit that matches, so completeness is a property of the operation rather than a hope about the cut-off.
- What happens if a document is still being indexed?
- The run parks and resumes when ingestion finishes, and the manifest shows the document as awaiting index throughout. It is never skipped, because a silently dropped document is the exact failure the feature exists to prevent.
- Does it find references that do not use the party's name?
- Defined terms and aliases are resolved when the document is ingested, so a reference to “the Purchaser” reaches the same entity as the party's name, and each mention records how it was resolved.
Works with
Tabular ReviewSchema-driven, spreadsheet-style extraction across thousands of documents with confidence and evidence.ExploreAssistantGrounded research, review, drafting and deep research across your sources.ExploreCitation FirewallThe wedge: every rendered output resolves to a citation on a source — or it isn't shown.Explore