Practice Note

Recorded synthetic worked example · reviewed 9 August 2026

Two sources, one checkable answer.

A recorded synthetic worked example shows how an AI chat assistant answers a maintenance question from five approved documents with two checkable sources—and stops when the evidence is insufficient.

Example type: Recorded and entirely syntheticEvidence date: 9 August 2026Content reviewed:
In this Practice Note
Example at a glance
Data
entirely fictional and synthetic
Documents
five one-page PDFs
Supported answer
two answer clauses · two citations
Unsupported cases
two · zero citations each
Model calls
none in this recording
External search or action
none
Human review
required

Important boundary: This is a predefined recording with entirely fictional documents—not a customer case, current live run, or deployed production system. The recording involved no model call, external search, or external business action. Human review remains required.

The problem

Maintenance information can exist and still be hard to use. People must know which document contains the answer and then find the exact passage themselves. A document assistant can shorten that search without replacing review by the responsible person.

Which documents were available?

The knowledge base contained five one-page PDF files. Every document belongs to the fictional Musterbetrieb Nordtal GmbH and is visibly marked as synthetic test data.

  • Maintenance plan 2026
    wartungsplan-2026.pdf
    Contains the next inspection date and responsible role.
  • Plant A17 data sheet
    anlage-a17-datenblatt.pdf
    Documents operating pressure, operating temperature, and the CIP limit.
  • Inspection report April 2026
    pruefbericht-2026-04.pdf
    Records the last documented inspection; the maintenance plan holds the next date.
  • Service agreement
    servicevereinbarung.pdf
    Defines the response window and explicitly keeps decisions with people.
  • Safety and preparation note
    sicherheits-und-vorbereitungshinweis.pdf
    Covers shutdown and depressurisation; the data sheet remains authoritative for limits.

How the supported question works

Question: “When is the next inspection of plant A17 scheduled, who coordinates it, and which documented operating limits must be considered during preparation?”

Answer: “The next inspection of plant A17 is scheduled for 15 October 2026; maintenance coordination is responsible. The data sheet documents a maximum operating temperature of 80 °C and maximum operating pressure of 6 bar; CIP cleaning is limited to 95 °C for 30 minutes.”

Source 1 · Schedule and responsibility

wartungsplan-2026.pdf, page 1, section “Prüfkalender 2026”: “Nächste Prüfung: 15. Oktober 2026 · Verantwortliche Rolle: Instandhaltungskoordination”.

Source 2 · Operating limits

anlage-a17-datenblatt.pdf, page 1, section “Betriebsgrenzen”: “Betriebstemperatur: maximal 80 °C · Betriebsdruck: maximal 6 bar · CIP-Reinigung: maximal 95 °C für 30 Minuten”.

The two distinct, relevant passages remain inspectable. The assistant does not approve, send, or execute anything.

What matters operationally is not the chat interface, but whether every used statement can be traced back to its exact source passage.

Recorded example in a light presentation with fictional documents · German interface · 48 seconds

The video shows a predefined example in a light presentation: five selected documents form the knowledge base. A team member asks one maintenance question that requires the schedule and the operating limits. The prepared answer shows two visible citations from two different documents, with page, section and exact excerpt.

English transcript and visual description
  1. 0 to 6s: The light presentation shows five documents clearly marked as fictional test data.
  2. 6 to 18s: One maintenance question receives two answer clauses: the inspection schedule and the documented operating limits.
  3. 18 to 27s: Two distinct citations open side by side: wartungsplan-2026.pdf and anlage-a17-datenblatt.pdf, each with page, section and exact excerpt.
  4. 27 to 35s: An unsupported energy question is reported as unanswerable, with zero citations.
  5. 35 to 41s: A question about another fictional company is refused without external search.
  6. 41 to 48s: The closing frame states the evidence boundary and confirms that people still review decisions.

What happens when the evidence is missing?

The same example contains two deliberately unanswered questions.

Information absent from the approved documents

Question: “How much energy did plant A17 use in July?”

Answer: “The approved synthetic document set contains no information about that. I cannot support a July energy-consumption answer.”

Citations: none.

Question outside the approved documents

Question: “What delivery time applies at the fictional Westhafen example business?”

Answer: “The question concerns another fictional business and falls outside the approved documents. This recording does not use external sources.”

Citations: none. The example neither invents an answer nor silently expands the agreed document boundary.

The principle in plain language

An answer is useful only when every statement can be traced to an approved source document. Here, one answer clause relies on the maintenance plan and the other on the data sheet. If no matching passage exists or the question concerns other documents, the answer is deliberately withheld. This recording involved neither a live search nor a model call.

What this example shows, measures and does not prove

Shown

The reviewed example shows five synthetic documents, two answer clauses with two matching citations, and two unanswered cases with zero citations each. Human review remains required.

Measured

No model-quality or business-performance metric was measured in this example. It is a predefined recording, separate from technical evaluations of document search.

Not proven

Live-model availability, correctness of arbitrary answers, customer performance, production readiness, time savings, economic benefit or a completed production deployment.

Could this approach fit your work?

That can be tested with a clearly bounded document set and representative questions. The important part is not the number of documents, but traceable sources, visible exceptions and one clearly responsible person.

Discuss a document assistant →