Skip to content

§Document retrieval

Ask questions of
your own documents.

We build retrieval systems that answer from material you already have. Every answer carries a citation to the passage it came from, and whoever reads it can open the source and check it in a few seconds.

Built forHealthcareLegalHigher educationFederal and defense

The problem

Your most useful documents are the ones you can’t paste into a chatbot.

The contracts, the protocols, the case files, the incident reports. People need to ask questions about them every day, and your obligations say they can’t be sent to a third party.

A general-purpose assistant has never seen any of them. Ask it anyway and it gives you a fluent, confident answer built from whatever it picked up in training. There is no way to tell that answer from a correct one by looking at it.

Both problems have the same fix. Keep the documents and the model inside your own network, and make every answer show where it came from.

What we do

We build one thing.

Document retrieval is the only service we offer. If what you need is close to it but not exactly it, ask.

Document retrieval

Ask your documents. Every answer cites its source.

A question-answering layer over documents you already have. Answers come back with citations, the quotations are lifted straight out of the file, and readers only ever see material they are cleared for.

  • Ingestion for PDF, DOCX and Markdown that tracks where each passage sits in the original file
  • Keyword and semantic search run together, with access control applied inside the search itself
  • Each claim checked against the source it cites before the answer reaches the reader
  • Conflicting and superseded versions flagged, so the answer never quietly picks one

Where it runs

Everything stays inside your boundary.

Documents, index, model, and answer all sit in infrastructure you already control. Nothing goes to a model vendor, so your reviewers have no new party to assess and no data-processing agreement to negotiate.

Where a private retrieval deployment sits relative to your network boundaryYour documents, the retrieval index, the language model, and the cited answer all sit inside your network boundary. The connection to external model vendor APIs is crossed out: no prompt, document, or answer leaves the boundary.Your networkYour documentsPDF · DOCX · MDRetrieval indexembeddings + lexicalLanguage modelrunning in your environmentCited answerverified claimsNo outbound callModel vendor APIsOpenAI · Google · Anthropic
Fig. 1A private retrieval deployment. The crossed-out line is the connection to outside model vendors, which this design does not have.

The system inherits the controls you already run: network segmentation, your identity provider, your logging, your change management. It is deployed inside them.

Proof

There is a working system you can look at.

A retrieval deployment over a set of policy documents, with verified citations, conflict detection, version tracking, and per-reader access control. It is running now, and we will show it to you.

Live system

“What approval do I need before making a purchase?”

Two policies are in effect. One sets the approval threshold at $500, the other at $1,000. Most assistants would pick one and state it confidently. This one reports the conflict, shows both sources, and declines to choose.

That matters because it puts a real inconsistency in front of the people who can fix it, where an ordinary answer would have buried it.

Abstains when the documents do not answer the question

Tracks versions so a retired policy cannot be quoted as if it were current

Records every verdict in an append-only ledger, where re-checking writes a new record

Times every stage and stores the trace alongside the answer

Why the answers hold

How the system checks its own answers.

These three properties come from how the pipeline is built, and they are the ones a reviewer will want to check.

01

A citation always points at a document that was actually retrieved

Each retrieved passage gets a temporary label that exists only for that request, and the model can cite only from that set. It never sees a real document identifier, so it has nothing to invent.

02

A quotation always matches the file it came from

The model is told not to quote. Quotations are cut from the stored document afterward, at the character positions recorded when the file was ingested.

03

A reader only retrieves what they are cleared for

The permission check runs inside the search, backed by row-level security in the database. If a code path forgets to set the reader’s scope, the database raises an error and returns nothing.

The claim checking is done by a second model, and that model stays fixed regardless of which model the reader picks for answers.

Security

Built for security review.

Your reviewers will ask where the data sits, what touches it, and how access is enforced. Here are the answers.

Where your data sits
In your tenant, under your identity provider and your logging
Model training
Nothing is sent to a model vendor, so there is no corpus for anyone to train on
Access control
Enforced inside the search and backed by row-level security in the database
Traceability
Citations limited to retrieved sources, quotations copied from stored positions, every verdict recorded in an append-only ledger

Who you would be working with

You work with the engineer directly.

The person who scopes the work is the person who writes it, and you will know their name before you sign anything.

For a security reviewer that is a simpler thing to assess. A named engineer with a public record of their work is a known quantity.

Next step

See it running.

We will show you the system answering real questions over a policy corpus, including the cases where it refuses to answer, and give you access to try it yourself. Bring the documents you have in mind and we will talk through how the same system would work on them.

Useful to know beforehand

  1. 01Roughly how many documents are in scope, and what formats they are in
  2. 02Who is allowed to read what, and how that is enforced today
  3. 03Any framework you are assessed against

Questions

What people ask first.

If your question isn’t here, send it over and we will answer it directly.

How is this different from Copilot or ChatGPT Enterprise?
Those are hosted services. Your prompts and documents are processed on the vendor’s servers under the vendor’s contract. If your obligation is that particular documents never leave your network, no contract term changes where the processing happens. We build for that situation: the documents and the model both stay on your side.
What document formats can you work with?
PDF, DOCX, and Markdown. During ingestion the system records where each passage sits in the original file, and that is what lets it quote a passage exactly later on.
Can it respect the permissions we already have?
Yes. The permission check runs inside the search, so a document the reader isn’t cleared for is never a candidate. Filtering results afterward would be too late, because the passage would already be in the prompt. Row-level security in the database backs the check up.
What happens when the model is wrong?
Plan on it. Before you see an answer, the system checks each claim against the source it cites and drops the ones that don’t hold up. When the documents don’t answer the question, it says so and stops.
How do we try it?
Ask for a demo. We’ll walk you through a running system answering questions over a set of policy documents, and you can have access to try it yourself. If you want to see it running on your own material, we can scope that on the same call.