'Ask the Record' Finds Nothing: 2 Reasons Why

An empty result on the screen; the boxes beside it may still hold the answer.

A blank answer is the most persuasive answer an AI tool can give. It asks nothing of you. It simply closes the door.

When you put a question to the record in a retrieval-augmented generation (RAG) tool, meaning a system that answers only from passages it first pulls out of your own collection, and it replies that it found nothing, one of two things has happened. Either the fact is genuinely absent from the collection, or retrieval failed to surface a passage that is sitting right there. The screen looks identical in both cases.

The technical explainers stop there. Litigators cannot.

An attorney who mistakes a retrieval miss for a true gap may certify a discovery response that is wrong, drop a deposition line that would have landed, or concede a point the evidence actually supports. A silence, misread, becomes a statement. And a statement signed by counsel carries weight long after the search that produced it has been forgotten.

So the useful question is never only whether the fact is there. It is also whether the search reached the place where the fact would live. I find that the second question is the one most often skipped, and the one a source-linked tool lets you answer in a single step, by reading what the answer cites before anyone pays to re-read the set.

What separates a genuine gap from a retrieval miss?

A genuine gap means the fact was never collected; a retrieval miss means it sits in the record unseen. Because every answer links to its exhibit, one look tells which.

Before you trust a blank, take three steps:

1. Ask a neighboring question, one step wider than the first.
2. Open the exhibits that answer cites, if any come back.
3. Judge whether those exhibits touch your subject at all.

A genuine gap is an absence in the collection itself. The document was never gathered, the custodian never imaged, the date range drawn too tight. No search, however patient, finds what was never there.

A retrieval miss is quieter. The fact lives in the record, but the system never carried it to the model. Retrieval is the step where a RAG tool (retrieval-augmented generation, in which the model answers only from passages it first pulls from your documents) chooses which pages to read. Choose badly, and the answer goes dark.

On the screen, both look the same. Silence, in a word.

The common assumption is that a blank answer is the cautious one, the answer least likely to mislead. I hold the opposite view. In discovery, a false absence misleads as surely as a false presence; it simply does so more politely.

The difference shows itself through the citations. In Relevant's design, every answer traces to its source: outputs link back to the exact exhibit they came from, so an attorney can one-click verify before filing. When the wider question returns exhibits from the same thread, the same custodian, the same week, the subject is plainly in the collection. The first blank was a miss. When the wider question also returns nothing near the subject, the light falls on empty ground, and the gap is real.

In practice, the citations become a lamp held at the edge of the blank. You see what surrounds the silence.

This check needs two things at once. Enterprise platforms have long supplied the defensible spine (Bates numbering, privilege handling, chain of custody); the newer AI entrants supply cited question-answering and chronologies. Our approach treats the pair as one tool, because the diagnosis needs both: the citation to see, and the custody record to trust what is seen. Phrasing the wider question well matters too, and the tradeoffs between keyword retrieval and plain-language questions are laid out in our comparison of concept search versus Ask the Record.

The check is also cheap. AI-assisted review runs cents per document, while re-reading the set by hand runs dollars per document. One wider question against a full re-read: the arithmetic is not subtle.

A comparison of 3 evidence items here, on source-linking, on combined tooling and on review economics, points one way. The diagnosis belongs inside the tool, not after it.

Yet some silences are made on purpose, by the firm itself, before the first question is ever asked.

Why do firms create their own false gaps?

A firm that holds privileged folders back from an AI tool manufactures its own gaps. Processing that runs single-tenant, under your own keys, with no vendor retention or training, removes the reason.

The diagnosis in the previous section works well. It rests, however, on one quiet premise: that the collection is whole.

Often it is not. And the hand that removed the pages belongs to the firm.

The fear is understandable. A partner hears that an AI vendor might retain uploaded files, or that a model might learn from them, and the instinct is to protect the most sensitive material by leaving it out. General counsel's email goes unloaded. The board folder stays on the shelf. The custodian whose messages touch legal advice is quietly dropped from the upload, and nobody writes the omission down.

Months later, someone asks the record a question whose answer lived in exactly that folder. The blank that comes back is not a retrieval miss, and it is not a true gap in the evidence. It is a gap in the firm's own upload, wearing the mask of an absence.

That third kind of silence is the most dangerous. It survives the neighboring-question check, because nothing near the subject was ever loaded. The citations return nothing. The attorney concludes the fact does not exist. Yet the document sits on a server down the hall, privileged perhaps, but real, and quite possibly responsive.

The remedy is to remove the reason for withholding rather than to manage the consequences. In Relevant's architecture, privileged evidence never leaves your control: processing runs single-tenant, or inside your own AWS account under your own keys, with no vendor retention and no model training. Nothing loaded trains anyone's model or waives privilege. Our view is that in-account deployment under the client's own keys is a trust posture no consumer AI tool and almost no small-firm tool can match, and it is the posture that lets a firm load the whole matter instead of a nervous fraction of it.

Loading everything does not mean producing everything. Privilege still has to be guarded at the door out. A fail-closed privilege gate on production, paired with immutable originals, content hashing and an append-only audit trail, keeps privileged material inside review while letting it inform answers. The mechanics of that separation are set out in our piece on designing a privilege gate for review.

So the order matters. Protect the exit. Open the entrance.

Fear closes doors. Sometimes it closes the very door the answer stood behind.

Sometimes a client still insists that one folder stay out. Then let the omission leave a mark. An append-only audit trail records what entered the set and when, so the withheld folder becomes a known edge rather than a hidden one.

The difference shows in a single sentence. Before: "The AI found no reference to the board's vote." After: "The AI found no reference to the board's vote in the loaded set; the board folder was withheld at the client's request and was not searched." The first sentence closes a question. The second leaves it honestly open, with a lamp set beside the door.

It also changes what you tell the client. A blank from a whole collection can be reported as a finding. A blank from a trimmed one can only be reported as a limit, and it should be, in those words.

In a word, an absence you recorded is a fact. An absence you forgot is a snare.

When the collection is whole, a blank can finally mean something, and the next task is to prove what it means to a client, to opposing counsel, and if need be to a judge.

How do you confirm a blank result without paying for a second review?

Confirm a blank with a short, logged checklist inside the same tool, not a second manual review. Built for a solo or small firm on one messy matter, the work needs no administrator.

A blank answer is a claim. Before it reaches a client letter or a discovery response, I want it to carry its own proof, the way a witness carries a name.

Four steps, in this order:

1. Confirm the collection is the one you think it is. Check that the expected custodians and the expected date range actually appear in the loaded set, and that the processing record shows the run finished.
2. Ask the wider question. Step back one level from the original question and note which exhibits the answer cites, or that it cites none.
3. Ask in the documents' own words. Use the names, nicknames, project code words and languages the custodians actually used, not the vocabulary of the complaint.
4. Write the finding down. Record the questions asked, the date, the scope searched and what came back, so the absence has a paper body.

Step three deserves a moment. Documents rarely speak the way pleadings do. A contract dispute may live in emails that never say "contract" at all, only a first name and a dollar figure. Where custodians wrote in more than one language, the vocabulary problem doubles, and our note on AI document review on foreign-language ESI covers what changes.

None of these steps asks anyone to re-read the set. That is where the savings live.

Conventional wisdom says that confirming a negative is the expensive part of review, the part that justifies a second pass by hand. I see it differently. A second pass is how firms confirmed negatives when the tool could not show its work. When each answer points to its exhibits, the confirmation becomes a few questions and a written note.

The structure matters as much as the cost. Enterprise platforms tend to price out exactly the firm this checklist is for, and they often cannot serve a single messy matter without an administrator standing between the attorney and the record. Enterprise-grade defensibility at small-matter economics is the design goal Relevant built toward, so that the attorney who asked the question is the one who can confirm the answer.

My advice is to test this before you rely on it. Bring a messy collection and a hard question, and watch the tool read, code and cite your own evidence rather than reading a case study about someone else's. Ask it something you already know is absent. Then ask it something you know is buried.

Watch both answers. The way a tool handles a true absence tells you whether its blanks can be trusted.

One more habit belongs beside the checklist: confirm the full answers as well as the empty ones. Because every answer links to the exact exhibit it came from, an attorney can open the source with one click before anything is filed. That link is the structural answer to the sanctions risk associated with Mata v. Avianca. Light falls on the page itself, not on a summary of it.

The two checks mirror each other. A blank asks what the search reached. A full answer asks whether the exhibit says this. Both are settled by the same link, read by the same attorney.

Keep the notes where the evidence lives. When Bates numbering, privilege calls and chain of custody sit in one tool with the questions and their citations, the record of your confirmation rests beside the documents it concerns, not in a separate memo that drifts away from them over the months of a case.

And the stakes of trusting a blank wrongly are not abstract; they arrive as a signature on a discovery response.

What should a blank answer mean to you from now on?

A blank answer should mean "check the citations," not "the fact is absent." Before any blank enters a client letter, attach the questions asked and the exhibits cited to the file.

I expect this to harden into practice. Courts and opposing counsel already ask how an affirmative answer was found; sooner or later they will ask the same of an absence. The firm that can show what it asked, what came back and what surrounded the silence will hold its position. The firm that can only say the tool found nothing will not.

Date the note, and keep it with the matter. The research section above walks through these findings.

If you are weighing platforms now, ask each vendor one plain thing: when your tool finds nothing, what does it show me next? The answer tells you more than any feature list, and it is worth hearing before your next certification is due.