VOXELL

Corpus Quality · document quality for AI answers

Your documents say refunds take 30 days. And 14.

Voxell answers your team’s and your customers’ questions from your own documents, with the source under every answer. When two documents disagree it puts that answer on hold before anyone is told the wrong one, and gives the document’s owner a one-click fix.

Score your documents

Sign in free, no card. Start with pre-scored examples, or add your own documents.

Your documents are never used to train models.

Granite Fork Helpwhat a customer sees · sample company
A customer asks“How long do I have to request a refund?”
  1. Without the check · a new support macro was imported that afternoon

    Refunds are available within 14 days of purchase.

    Source: Macro: Refund request (v2)

    Refund policy §2 says 30 days. Neither document says when it applies, so search can return either, and each sounds just as sure.

  2. With Voxell · the answer goes on hold as the macro arrives

    For the current refund window, please contact our support team.

    Two documents disagree, so no number is given. The document’s owner gets both sentences side by side.

  3. One click later · the owner picks 30 days

    Customers can request a full refund within 30 days of purchase.

    Source: Refund policy §2

Import checks run as each document arrives, and are on by default. The report further down shows a day the sample company had them switched off, and what that cost.

Your first ten minutes, on your own

  1. 1

    Paste two documents you suspect disagree

    Or drop in text or Markdown files: Notion and Google Docs both save a page that way from their export menu. Free, no card.

  2. 2

    See the contradiction caught

    Both sentences side by side, with a button for each. Every document also gets a grade and its weakest passages, worst first.

  3. 3

    Share your Ask page

    A link your team opens without accounts. Every answer shows the passage it came from, or says it does not know.

PDFs, Word files, a Drive folder or a Confluence export? Send them through the free scan form: we load up to 500 documents for you and send back the grade, every contradiction and a report within two business days.

An answer is only as right as the passage it was written from.

To answer a question, search picks a few passages from your documents and the AI writes from them. If one is out of date, a duplicate or half a table row, the answer is wrong and sounds just as sure. Even purpose-built legal AI research tools do it (Stanford RegLab study, 2025). In a company’s documents it looks like this; the examples are from Granite Fork, the sample company used on this page:

Facts that contradict each other

Two prices, no dates

$18 per seatPricing FAQ$22 per seatQ3 price sheet

A new price sheet arrived without a date. That afternoon 9 answers gave sales staff one price or the other, each sounding equally sure.

Duplicates

The same sign-off, 380 times

Thanks for reaching out! We’re sorry to hear you’re having trouble. Our team is here to help. Please let us know if there is anything else we can do.

It matches every question and answers none, pushing the real answer down the results.

Broken passages

Half a table row

| NW-2 | Tacoma | 06:00–

Half a table row. The column names (Region, Depot, Hours) are in a different passage.

The page your team asks on

This is where people ask: an Ask page is a collection published as a page with an Ask box, which your team or your customers open from a link. (In the product and the API it is called a Space.)

  • Every answer shows the passage it came from, so anyone can check it in ten seconds
  • Where two documents disagree, it says so instead of picking one
  • Where your documents do not cover a question, it says that instead of guessing
  • What people are told while an answer is on hold is yours to word
  • Share it by link with no accounts, or limit it to people who sign in with your company’s email

Your first Ask page is free, with 50 answers a month. Your own name and look on the page come with the $100 plan and with Team.

Granite Fork Supportan Ask page · sample company
Ask about billing, refunds or your accountAsk

How do I reset my password?

Open Settings, choose Security, and select Reset password. A reset link is emailed to you.

Source: Account security §2 · open the passage

How long do I have to request a refund?

For the current refund window, please contact our support team.

Two documents disagree, so no number is given until the owner picks one

Do you offer a student discount?

I don’t have enough information in these documents to answer that.

No guess: the documents do not cover it

How it fits with your documents

1

Voxell keeps its own copy, and you keep it current

One collection per team or topic. Text and Markdown go in directly; PDFs, Word files and whole exports go through the free scan. Your original files stay where they are, untouched.

There is no automatic sync yet. When a document changes, paste or drop in the new text under the same name (or have a script do it through the API): each upload is a new version, checked as it arrives. A changed PDF or Word file is saved as text first, or sent through the scan form again.

2

Each collection is measured two ways

A document grade from 0 to 100: is each passage clear, specific and complete enough to answer from? And a set of 20 to 60 test questions, drafted from your documents and from real searches. The collection’s owner, the person who looks after those documents (probably you), approves each question from an emailed link; nothing counts until they do.

3

Fixes are yours to approve, and to undo

Voxell applies a fix once the owner approves it. It changes the copy Voxell searches and keeps every earlier version, so it rolls back in one click. Your original files are never edited; where an original needs the same change, you get the exact sentence. A later upload of the same document keeps the fix unless it changes that passage, in which case you are asked.

Find it, fix it, prove it

1

Find it, in every collection

Every collection gets a grade and a count of test questions answered from the right passage. Facts that contradict each other come first, and each is put on hold as it arrives: no answer states either value as fact until its owner picks. Customers are sent to support; staff are shown both, marked as disagreeing.

FLegal templatesLegal · 6 of 20 questions found36/100
FRegional operations tablesField Ops · 7 of 20 questions found41/100
DSales price bookSales · 12 of 20 questions found · 9 conflicting facts64/100
CField support handbookField Ops · 18 of 25 questions found · 3 conflicting facts67/100
2

Fix these pieces

The weakest passages, worst first, with the reason in plain words. Fixes are grouped by cause and tried on a private copy before anything changes. The collection’s owner approves with one click, and can undo it the same way.

18/100
Depot hours (North-west)
| NW-2 | Tacoma | 06:00–
Re-split tables by whole rows, column names attached7 → 13 of 20 questions found
grade 41 → 68Tested on a copy
3

Prove it worked

Search is measured again on the same questions, approved by your own people. Every change, including new content that set you back, shows what it moved.

Test questions found, all 14 collections204 → 236 of 323
Real searches answered first time: no second search within a minute69% → 75%
Healthy collections: grade B or better, 80% of test questions found, no facts on hold2 → 4 of 14
Contradictions settled for good, by giving each value its dates18

The method

A systemic approach to achieving measured retrieval.

No more “search feels off”. Every collection is checked the same way, and every claim on this page is a number you can run again.

Systemic

Every collection and every import, not one fix at a time. Problems are fixed by cause across all of them, and new documents are checked on the way in, so a fix stays fixed.

Measured

Two computed numbers, before and after every change. On the document side, a grade from 0 to 100 computed from the passages themselves (in the API it is called VQS, the Vector Quality Score). On the search side, how many test questions, approved by your own people, find their answer, and how high it ranks (nDCG@10, where 1.00 means the right passage always comes first: across the sample company, 0.58 → 0.65). A third number checks the first two from real use: how often a search is answered first time, meaning nobody had to search again within a minute (across the sample company, 69% → 75%).

Stop problems coming back

Import checks are on by default: as new documents arrive they get each collection’s saved fixes, and a document that contradicts an existing one has its answers put on hold for its owner as it arrives. A nightly check sweeps up anything that got in another way. The sample company switched import checks off for a bulk load on Jul 1 and did not switch them back on, so every setback this quarter came in with an import:

  • Sep 12 180 section headings stored apart from their text
  • Sep 21 2 conflicting safety settings, 1 other conflict and 120 headings stored apart from their text
  • Sep 25 9 conflicting prices and terms
  • Oct 5 1,140 near-duplicates, 380 copies of a canned sign-off and 5 conflicting facts

The report you hand your manager

A sample quarter, with its bad day left in: what changed, what each change moved, what set you back, and what customers were told in the meantime. It names teams, never people, and where it relies on someone’s word it says whose. The sample is the Support team’s, two days after the import at the top of this page. Their admin had switched import checks off; the report shows what that cost and who turned them back on.

  • One page to send, as a PDF or as a link only your company can open; a sent copy does not change afterwards
  • The same report for one collection, one team or the whole company
  • Every figure adds up to the rows beneath it

Search health report · Support knowledge base Jul 1 to Oct 7 · as of Oct 7, 15:02

Search finds the right passage for 51 of 60 test questions, up from 37. In real use, 76% of searches were answered first time in the week before the import, up from 66% in July. 5 fixes did the work; 1 import of new content set it back.

That import, on Oct 5, brought 5 facts that contradicted an existing document. Answers on them went on hold that night; in the 9½ hours before, 44 answers went out with one value or the other, 31 of them to customers. The owner settled all 5 on Oct 7: 19 of the 44 were wrong, 14 to customers.

The largest: 11 said the refund window is 14 days when it is 30 days.

Of the 19 wrong answers, 10 can be traced to a person (5 signed-in customers, 5 staff): the owner reports them told. The other 9 went to customers who were not signed in. The owner also reports the 5 source documents corrected.

How it got through: a workspace admin had switched import checks off on Jul 1 for a bulk load and they were not switched back on. The owner turned them back on for this collection on Oct 7.

Test questions answered from the right passage37 → 51 of 60
Real searches answered first time66% → 76%
Document grade58 → 77 of 100
Facts that contradict each other, still on hold0 of 5
Engineering work neededNone
What changedTest questions found
Removed exact duplicatesJul 14+5
Kept each heading together with its textJul 28+2
Left sign-offs and disclaimers out of searchAug 25+6
Support macros imported from the helpdeskOct 5 · new content-2
Merged near-duplicate passagesOct 7+3
Left the new macros’ canned sign-off out of searchOct 7±0
Total since Jul 1+14

Is our data safe?

Your documents are never used to train models. They are encrypted in transit and at rest, every company that processes them is named, and you can delete any of them at any time.

Security and buying, on one page → · For developers: the calls, with real responses →

Start with your own documents

Your first collection

Freeno card

One collection, up to 500 documents: an Ask page you can share, with the source under every answer (50 answers a month), the checks on this page, and the report. Put text in yourself today, or send PDFs, Word files and exports and we load them within two business days.

Score your documents or send PDFs and exports for a free scan →

Enough answers for a team

$20/month, cancel anytime

The same Ask page with 10,000 answers a month instead of 50, and the checks still on that first collection. Your team asks from the link without accounts. Only you pay.

Every collection, your branding

$100/month, cancel anytime

The checks, grades and one-click fixes on every collection you own, 50,000 answers a month, and your own name and look on the Ask page. Still one person, on your own card.

A team, from 5 people

$80/person/month, billed yearly

That price is for the people who check and fix documents; people who only build with the API are $15, and people who only ask questions or approve a fix cost nothing. One bill and central admin.

Cost at your size →

Compare all plans → · Building it into a product? The same prices apply to the API. Limits and price for developers →

Voxell is a small team in California and a member of NVIDIA Inception. MTEB is the public benchmark for how well a model finds the right text: our research model, Ingot-8B-R3, ranks #1 for English on the public MTEB leaderboard (English v2).