Chunkless
§ 00 · A research engine for private documents

It reads your whole library the way its author would.1

Ask a question and Chunkless finds the clause in one file, the exception in another and the memo that quietly overrode both. It answers with the exact section and page, checks every statement against its source, and says plainly when your documents disagree or don't say.

Free during early access · no waiting: sample libraries are already loaded.

1 On 34 questions that each need several documents, half held out from tuning: 0.79 of the required facts, against 0.59 for standard chunk search. Paper, §5

§ Coverage, held out
0.79
of the facts a complete answer needs, on questions that span many documents
§ Chunk search, same test
0.59
the standard chunk-and-retrieve pipeline, same questions, same blind judge
§ Evidence read
99%
of the evidence a complete answer needs is in what it reads, across 32 documents
§ Statements checked
100%
of the statements shown were checked word for word against the cited source
§ 01 · Why "Chunkless"

Most AI search cuts your documents into chunks and grabs the few that sound like your question. But your answer lives across three documents. Chunkless keeps every document whole and follows what it points to, the way the person who wrote it would.

Scroll · the chunks become one document
Chunk-and-hope search
  • Returns the passages that sound most like the question
  • Treats a 2019 policy and its 2024 replacement as equals
  • Fills gaps with plausible prose, and skips the diagram that holds the answer
Chunkless
  • Follows references: a definition in one file, a limit in another, the memo that changed it
  • Knows what overrides what, and answers with the rule that's actually in force
  • Checks every statement against its source, reads every figure, and names what's missing
§ 02 · What it does

Built for the questions no single document can answer.

Fig. 3 · Your library, as one structuredepartments rise · links are drawn · a question travels them
§ 2.1

Links anything to anything

It learns your documents' vocabulary: the same thing under three names, what refers to what, which version replaced which, who cites whom. Then it connects volume 1 to volume 20 on its own.

versions · citations · continuations
§ 2.2

Answers with the page

Every part of the answer traces to its section and page. Open the source, see the passage, check it yourself.

section · page · figure
§ 2.3

Says what it can't prove

Each answer says whether the documents settle it, settle it only under a condition, point elsewhere, or contradict each other.

four verdicts, below
§ 2.4

Explains why things break

Ask "why does this keep failing?" and it follows the chain through runbooks, configs and procedures to the step that causes it.

root cause
§ 2.5

Reads what search can't

Figures and screenshots are described at upload, tables keep their rows and columns, scans are OCR'd. A manual answers with each step's picture.

figures · tables · scans
§ 2.6

Thinks with you

Give it the papers or specs and ask for a design, a theory or a plan. It builds the answer from what the documents prove and marks what is still open.

research mode
§ 2.7

Private by construction

Each organization is isolated at the database's row level. Connector keys are encrypted at rest. Organize by department, folder and case, and bring your own model key if you like.

isolation · encryption
§ 03 · Honest by design

Every answer comes with a verdict.

A good expert tells you how sure they are and why. Chunkless works the verdict out from your documents instead of guessing it.

CompleteCOMPLETE

The documents settle it.

Everything the answer needs is in what you uploaded, and each piece is cited.

"Leave carries over 5 days." HR Handbook §6.2
ConditionalCONDITIONAL

True, if…

The answer depends on something your question didn't say, and it tells you exactly what.

Yes, if the vendor is already on the approved list (§2.4).
MissingMISSING

A document points elsewhere.

The rule refers to a schedule, form or policy that isn't in the library yet, and it names it.

§9 depends on "Schedule B", which isn't uploaded.
ConflictingCONFLICTING

Two sources disagree.

Both are in force and can't both be right. It shows you both, and who decides.

Memo 2025-07 vs Policy §4.2: the stricter applies until Finance rules.
§ 04 · How it works

Heavy reading once. Sharp answers after that.

It reads everything carefully up front, remembers how it all fits together, and when you ask, it goes back to the sources and checks before it speaks.

Fig. 4 · One document, taken apartwhat Chunkless keeps when it reads
  1. Cover and title
  2. Section tree
  3. Tables, kept as grids
  4. Figures, described
  5. Facts, with their conditions
  6. Links to other documents
4.1

Bring the library

Upload files or point it at a bucket. Group them in departments and cases.

PDF, Word, slides, sheets, scans · up to 1 GB
4.2

It reads everything

Sections, tables and figures stay in place, with every rule, limit, definition and cross-reference noted.

every figure described · what overrides what
4.3

You ask anything

A quick lookup, a "can I…?", a "why does this keep breaking?", or a design built from thirty papers.

one department or all · follow-ups keep context
4.4

It investigates

It opens the sources, looks for evidence against its first answer and checks every statement before you see it.

section and page · the trail, on request
Four ways to ask

Match the effort to the question. You pick, or it picks for you.

Quick

A fact, a number, a definition, in seconds.

Full

Reads until each part is found or shown absent, then checks every statement.

Auto

Quick for a one-fact lookup, full for everything else.

Research

Long-running reasoning: designs, theories, plans.

§ 05 · Who it's for

Anyone who inherits a pile of documents and a deadline.

Operations · HR · Finance · Legal
Can I approve this, and who says so?

A new hire gets the answer a ten-year veteran would give, with the policy, the memo that changed it and the exception that applies.

Engineering · support · IT
Why does the export fail every Monday?

It follows the chain through runbooks, configs and procedures to the step that's actually causing it, and shows the chain.

Research · product · strategy
From these papers, what should we build?

It pieces a design together from what each source proves, says where they conflict and marks what's still open.

§ 06 · Published research

Beyond Chunks: answering questions across a document library by reading whole sections and checking every claim

We publish how it works and how it scores, including what didn't help and where we corrected our own numbers.

Read the paper →
Bonthu & Kamatham · Grounded IQ · 2026 · doi:10.5281/zenodo.23403889
Fig. 2 · held-out coverage, 17 questions
Chunk search0.59
Chunk search, reranked0.65
Chunkless, everyday0.79
Chunkless, research0.81
Whole library in one call0.95
The last bar is an upper bound that only works when the whole library fits in one model call.
§ 07 · Early access

Try it now, free while we're in early access.

This is a working build, not a demo video. Make an account and you're in: sample client files and their libraries in tax, insurance, banking and medical are already loaded, so you can ask your first question in a minute. Then bring your own documents. You get a direct line to the people building it.

  • CostFree during early access.
  • Sign-upName, email and a password. Nothing to wait for.
  • Ready to tryShared sample libraries (fictional clients and public reference texts), read-only, beside a private workspace of your own.
  • Your access7 days from the day you join. Genuine feedback through the Feedback button in the app earns extra days.
  • Your dataWhat you upload goes to your own workspace, walled off from every other tester. Please don't upload anything you aren't allowed to share with a test system.
  • What we askTry it on real questions and tell us where it's wrong. That feedback shapes what ships next.

Start free

Account · 1 page
We use your email only for your account. No newsletter, no sharing.
ON ITS WAY

Check your inbox.

If a seat is open, your personal invite is on its way now: one click and you set a password. It only works with the address you gave us. If every seat is taken, you're on the list and we'll email you when one opens.

Nothing arrived after a few minutes? Check your spam folder for mail from Chunkless.

§ 08 · Questions

Before you ask.

How is this different from ChatGPT with file upload?

Those tools read a handful of passages that sound like your question. Chunkless reads your whole library up front and learns what refers to what, which rule overrides which and where each figure belongs. When you ask, it investigates and checks every statement against its source. So it can say "these two documents disagree" or "the answer depends on a schedule you haven't uploaded" instead of writing something that merely sounds right.

How does sign-up work?

Your name, your email and a password, and you're in: the shared sample libraries (fictional client files with the tax, insurance, banking and medical reference texts) are already loaded, read-only, beside an empty workspace of your own for 7 days. While seats last; when every seat is taken, you're on the list and we email you when one opens. The Feedback button at the top of the app reaches us directly, and genuine feedback earns extra days.

Can I try it before signing up?

Yes, two ways. "Try it on your own documents" opens a private workspace with no account: drop in up to 3 files and ask 5 questions. Only you can see it, and everything in it (the files, what was read from them and your questions) is deleted after 24 hours. Or "try a sample library": eleven documents from a fictional medical-device company that you can browse and ask about.

Where do my documents go?

They're stored encrypted on AWS in the US. Each organization is isolated at the database's row level, so no other tester can see your documents. To answer questions, the engine sends the relevant passages to a language model (OpenAI by default), or you can bring your own model key. During early access, please only upload documents you're allowed to share with a test system.

What kind of documents work best?

Collections where the answer is spread out: policies and their amendments, SOPs and manuals, contracts and schedules, research papers, clinical protocols, engineering runbooks. It handles PDFs (including scans), Word, PowerPoint and Excel.

Why is access limited?

Reading documents deeply takes real compute. We'd rather give a small group a fast, dependable engine than give everyone a slow one. We open it to more people as we scale.