Skip to content

§Walkthrough

Five things a retrieval system should do.

These are transcripts from a running deployment answering questions over a corpus of twenty policy documents. Four of the five you can reproduce yourself. The fifth is the most important and cannot be shown in a shared demo, which is why it is written down here.

The live deployment sits behind a shared credential rather than a public link, so we send access out on request.

01

A cited answer

Verified

> How many days do I have to submit an expense report?

Every claim below was checked against the source it cites.

Expense reports must be submitted within 30 days of the expense being incurred. [1]

References

  1. [1]

    Expense Reimbursement Policy

    v2.1 · effective 2024-03-01

    Employees shall submit all expense reports within thirty (30) days of the date the expense was incurred.

What to notice

The quotation is not written by the model. It is sliced out of the stored document at the character offsets recorded during ingestion, so it cannot drift from what the file says.

02

A conflict it refuses to resolve

Policies disagree

> What approval do I need before making a purchase?

More than one policy is currently in effect and they do not agree. Both are shown.

Purchases require manager approval above a threshold, but the two policies in effect state different thresholds. [1] [2]

References

  1. [1]

    Expense Reimbursement Policy

    v2.1 · in effect

    Purchases exceeding five hundred dollars ($500) require prior written approval.
  2. [2]

    Procurement and Purchasing Policy

    v1.3 · in effect

    Any purchase over one thousand dollars ($1,000) must be approved in advance.

What to notice

This is the scenario worth watching. A confident assistant picks one number. This one reports that the documents contradict each other and declines to choose, which sends a real problem back to the people who can fix it.

03

A superseded version, caught

Verified

> How long after travel must I submit my travel expenses?

Every claim below was checked against the source it cites.

Travel expenses must be submitted within 30 days of the end of travel. [1]

References

  1. [1]

    Employee Travel Policy

    v4.2 · current

    Travel expense reports are due within thirty (30) days of the final day of travel.
  2. [2]

    Employee Travel Policy (superseded)

    v3.0 · superseded

    Reports are due within forty-five (45) days of return.

What to notice

Both versions are in the index. The retired one is labeled rather than deleted, so the answer is current but the history is still auditable.

04

A question it will not guess at

No answer in the corpus

> What is the policy on keeping tropical fish in the office?

Nothing in the policy corpus available to you answers this question.

No answer could be supported by the policy corpus for this question.

What to notice

Saying nothing is a feature, and it is surprisingly rare. A system that always produces something teaches its readers to stop checking, which is the failure the whole design is meant to prevent.

05

The same question, two readers

Depends who is asking

> What are the compensation bands, and what does band 4 pay?

Asked by a reader outside the HR group, this returns the same refusal as a question about tropical fish.

Reader in the HR group: the bands and the band 4 range, cited to the compensation policy. Reader outside it: “No answer could be supported by the policy corpus for this question.”

What to notice

The two refusals are byte-identical, so the system cannot be used to work out that a restricted document exists. This is the strongest property in the build and it is the one you cannot see in the public deployment, where everyone arrives as the same principal.

Provenance

What is ours, and what is not.

The system is built to attribute everything it says. The same standard applies to how we describe it.

Ours

The retrieval engine: ingestion and chunking, the hybrid retrieval and fusion, the authorization model and its row-level security, schema-constrained generation, offset quotation, independent claim verification, conflict and supersession detection, the release gate, and the append-only audit ledger.

Not ours

The chat interface is Atlas UI 3, an open-source project released by Sandia National Laboratories under the MIT license. We integrated our engine into it rather than rebuilding a chat client.