Skip to content

Engineering teams

The wiki, the datasheets, and the truth

An industrial electronics team pointed Cogeto at everything: the wiki nobody trusts, a family of forty near identical datasheets, and the decision threads. Now the corpus answers with citations, and the drift has names.

The situation

Documentation drift, unmeasured

The wiki nobody trusts

Three people have edited it since 2022. Everyone suspects it is wrong somewhere; nobody knows where, so everybody re-asks a colleague instead.

Forty datasheets, shared boilerplate

Near identical variants differing in numbers nobody can name from memory. Copy paste keeps the boilerplate consistent and the differences invisible.

Onboarding by asking around

Decisions live in threads and heads. A new engineer's first month is spent finding out which document to distrust.

The run

The run

One bulk import: the wiki export, the datasheet family, design notes, and the archived decision threads.

  1. 1486 documents imported in one operation
  2. 2Datasheets anchored per variant, so shared boilerplate stays distinct
  3. 3Wiki pages and threads read as sources with dates
  4. 4Verification against source sentences before storage
  5. 5Overnight reconciliation, variant against variant, wiki against datasheet

The findings

The findings

Three of the thirty three, as the product reports them.

The wiki against the datasheet: the drift everyone suspected, named.

Variant against variant: anchoring keeps forty datasheets distinct, so a real difference is a finding, not noise.

The signed report the team files with the quarterly review.

Then you ask.

Which datasheets still show the old sleep current? Two variants still carry 120 microamps: K-241 r2 and K-244 r1. The current value is 85 microamps since r3. Stands on: K-241 r2, K-244 r1, K-240 r3.

Is the enclosure IP54 or IP65? IP65. Conflict: the wiki hardware page still says IP54, flagged and linked. Stands on: Datasheet r7, Wiki page.

What it means for you

What it means for you

A knowledge base that proves itself

Every answer cites its sentence and document, or says the corpus is silent. Trust stops being a feeling.

Drift caught the day it happens

A wiki edit that contradicts a datasheet is a same day finding, not a legend that spreads for a year.

Onboarding from the corpus

New engineers ask the corpus and get citations, instead of spending a month learning which documents to distrust.

Variants kept honest

Forty near identical datasheets stay distinct per variant, so the sales engineer quoting one model never gets another model's numbers.

Run it where your material lives

  • Hosted in the EU: your single tenant instance, operated for you.
  • Your cloud: deployed inside your own infrastructure.
  • Fully offline: local models, no outbound connectivity, no telemetry.

First questions

Our documentation is honestly a mess. Is that a problem?

It is the point. The messier the corpus, the more the verification gate and the suppression log matter: what cannot be verified is logged, not believed.

Who fixes the findings?

You do, with both sentences in front of you. Findings are surfaced on the source, in answers, and in the report; there is no queue to babysit.

Does the wiki outrank the datasheet, or the reverse?

Neither. A conflict shows both sides with dates and revisions. A person decides; your confirmation then outranks the machine's judgment.

Start a pilot on your documents.

Send us the wiki export and the datasheet folder; we deliver the signed findings report.