Engineering Knowledge Assistant

Why did we design the 1998 series that way? Ask the works — and get an answer with sources, in seconds.

A transformer or generator works runs on two kinds of knowledge. The first is written down: calculation reports, design reviews, test-field protocols, works standards, decades of order documentation in Teamcenter and on network drives. The second lives in the heads of your senior engineers — the reasons behind the rules, the failures that shaped them, the shortcuts that are safe and the ones that are not.

Both are in trouble. The written knowledge is technically “available” and practically unfindable — engineers spend a quarter of their working day searching for information that exists. And the heads are retiring: the same years in which you are hiring more new engineers than ever, to build more units than ever.

The Engineering Knowledge Assistant is a works-specific AI assistant over exactly this knowledge: it reads your document base, respects your permissions, and answers engineering questions with the source attached — in the language your works actually speaks. On your infrastructure, not in a public cloud.

// The knowledge problem, in numbers

of an engineer's day goes into searching for information that already exists

0 yrs

of design decisions, calc reports and test protocols — findable only by the people who wrote them

0 mo

typical onboarding time before a new design engineer is productive on your works standards

seconds

to a sourced answer once the works' knowledge is behind an assistant instead of a folder tree

Your works has answered this question before. It just can’t remember where.

Every experienced engineer knows the moment: a young colleague asks why the winding exit is routed this particular way, or which clearance applies for a 245 kV bushing under your works standard — and the honest answer is “ask Herbert, he did the 2003 series.” Herbert retires next spring.

The documentation exists. The design review from 2003 is a PDF on a share, the calc report is a dataset in Teamcenter, the corrected clearance table is revision D of a works standard that revision C contradicts. Search finds none of it, because file names are order numbers and folder logic is tribal knowledge. So questions get answered the expensive way — by interrupting the three people who know, until they leave.

A knowledge assistant changes the mechanics: instead of knowing where to look, engineers only need to know what to ask. The assistant retrieves the relevant passages across all connected sources, formulates an answer, and — this is the part that makes it engineering-grade — cites its sources, so the engineer verifies in one click instead of taking a black box at its word.

// Illustrative example: assistant on the works' own knowledge base

Which oil temperature rise limit applies for ONAN distribution units under our works standard — and did we ever deviate?

Per the works standard, rev F (2019) the works limit is 55 K top-oil rise for ONAN distribution units — 5 K below the IEC 60076-2 limit, introduced after repeated test-field findings on the 1600 kVA series. Deviations were approved twice: one order in 2011 (customer spec, 60 K, with upgraded radiators) and one in 2016. Both deviation reports are linked below. Works standard rev F · p.12Test report 2018Deviation report 2011

New here — how do I set up the drawing frame for a customer in NX so the title block fills itself?

Use the works template, don't build frames manually: File → New → works customer-drawing template. The title block fills from Teamcenter attributes — set customer and order on the item first, otherwise the fields stay empty. The 2024 onboarding guide covers it step by step; note the known pitfall with legacy items created before 2020. CAD guideline · §3.2Onboarding guide 2024 · p.18Known issue: pre-2020 items

How the assistant gets its knowledge — and keeps your permissions

Under the hood this is retrieval-augmented generation (RAG): the assistant never “learns” your documents into a model that could leak them — it retrieves the relevant passages at question time, from an index that respects who is allowed to see what. The pipeline is straightforward, and every stage runs where your data lives:

01 — CONNECT

Your sources, as they are

Teamcenter items & datasets
Network drives & archives
Works standards & CAD guidelines
Wikis, protocols, test reports

02 — INDEX

Understood, not just stored

Documents parsed incl. tables & scans
Revisions resolved — latest wins, history kept
Permissions mirrored from source systems
Updates flow in automatically

03 — ASK

Where engineers already work

Side panel next to NX
Browser & intranet
Teams, if you want it there
German, English — your works' terms

04 — ANSWER

Sourced, or not at all

Every claim cites its document
One click to the original passage
"I don't know" instead of invention
Feedback loop improves retrieval

Four jobs the assistant does from day one

A knowledge assistant is not a toy for curious questions — it earns its keep in four recurring situations that every energy equipment works knows:

Onboarding that doesn't burn seniors

New engineers ask the assistant the hundred small questions — template locations, works conventions, "how do we do X here" — instead of interrupting the two colleagues who already carry the department. Months to productivity become weeks.

Works standards & norms, resolved

Which limit applies, which revision is current, where we are stricter than IEC and why — answered with the standard cited, instead of guessed from an outdated printout in a desk drawer.

Design history on demand

"Have we solved this before? What went wrong last time?" The assistant surfaces the 2009 design review and the test-field finding that explains today's rule — the why behind the works standard, preserved past retirement.

Stuck-in-NX help, works edition

NX questions answered in the context of your setup — your templates, your Teamcenter integration, your naming rules — where generic forums and even Siemens' own copilot can only answer generically.

When does a knowledge assistant pay off?

The value scales with two things: how much written knowledge you have, and how fast the heads that can navigate it are leaving. You are a candidate if most of these apply:

Decades of calc reports, design reviews and test protocols exist — in Teamcenter, on shares, in binders — but only a few people can find anything in them.
Several of your most experienced engineers retire within five years, while hiring is ramping up.
Seniors lose hours per week to interruptions — answering the same onboarding and works-standard questions again and again.
Works standards have revisions and exceptions that regularly cause confusion — or worse, wrong assumptions in designs.
Public AI tools are blocked or useless for these questions, because the answers live in confidential works documents.
You want the assistant on your infrastructure, permission-aware, connected to Teamcenter — not another cloud silo.

What the assistant connects to

Built around the sources an energy equipment works actually has — with permissions mirrored, not bypassed.

Teamcenter items & datasets Works standards & CAD guidelines Calculation & test-field reports Design reviews & lessons learned Network drives & PDF archives Wikis & intranet pages NX / NXOpen documentation SSO & permission mirroring On-prem / EU cloud LLMs

Before and after: a new engineer’s first month

The clearest way to see the assistant’s value is through the person who costs and suffers the most in the status quo — the newly hired design engineer:

Before — folder tree & hallway

Every question costs a senior's time — or stays unasked
Works standards read like law texts with invisible exceptions
Old design decisions look arbitrary — so they get "improved" and repeated mistakes return
Productive after ~6 months, confident after a year

After — assistant on the works' memory

Routine questions answered in seconds, with the source to verify
Standards come with their current revision and known exceptions attached
The "why" behind rules is one question away — history stops repeating
Seniors are asked what only seniors can answer — judgement, not lookups

Why inteliScience

Every IT consultancy can deploy a RAG stack. The difference shows in the details that decide whether engineers actually use it: does the assistant understand that works standard rev F supersedes rev E but the 2011 deviation still matters? Does it read Teamcenter datasets, not just file shares? Does it know what a top-oil rise is — and that inventing one would be worse than admitting ignorance?

We build engineering AI on the Siemens stack for a living. The assistant connects to Teamcenter through supported interfaces, sits next to NX where your designers work, and is tuned on your document types — calc reports, test protocols, works standards — not on generic office files. And where Siemens’ own Teamcenter Copilot provides useful building blocks, we integrate rather than reinvent; where it stays generic, we build the works-specific layer it cannot offer.

Sovereignty is non-negotiable: the assistant runs on-premises or in an EU cloud under your control, with permissions mirrored from your source systems. Your works’ memory stays your works’ property.

For quotation work the same source-linked principle applies: our AI similar project search takes a new enquiry and finds the technically similar projects your works has already delivered.

Fixed price — no day rates, no surprises

Knowledge Assistant Pilot

€14,500 fixed

2–3 knowledge sources connected (e.g. works standards, one Teamcenter area, one archive share)
Assistant deployed on your infrastructure — on-prem or EU cloud
Evaluation on 100 real questions, collected from your engineers
Measured answer-quality report — reviewed together with your team
Rollout blueprint: sources, permissions, NX-side integration, costs

Start the pilot conversation

6–8 weeks · your data never leaves your control · pilot results are yours, whatever you decide afterwards.

For comparison

The cost of searching

10 engineers × a quarter of their day spent hunting information
Senior hours burned on interruptions instead of design work
And the un-priced risk: the next retirement takes the answers with it

Frequently asked questions

What stops it from inventing answers?

Three mechanisms, in this order: the assistant may only answer from retrieved passages, never from the model's general memory; every claim must carry a citation the engineer can open; and when retrieval finds nothing solid, the honest answer is "not found in the connected sources" — with the closest matches listed. In the pilot we measure exactly this on 100 real questions from your engineers, so you see the grounded-answer rate before rollout, not after.

Do our documents leave the company?

No. Index, retrieval and language model all run on your infrastructure — on-premises or in an EU cloud under your control. Nothing is used to train any external model. For energy equipment works with critical-infrastructure customers this is usually a hard requirement, and the architecture treats it as one.

How are permissions handled? Not everyone may see everything.

Permissions are mirrored from the source systems, per user, at question time: the assistant only retrieves from documents the asking engineer could open themselves in Teamcenter or on the share. No shadow copies with relaxed rights, no "the AI saw it so everyone can ask it" loophole. Login runs through your SSO.

Isn't this what Siemens' Teamcenter Copilot does?

Partly — and where it fits, we build on it rather than against it. Teamcenter Copilot brings solid building blocks for chatting with Teamcenter-managed documents. What it does not cover: sources outside Teamcenter (shares, wikis, binders' worth of scanned reports), works-specific tuning on your document types and terminology, the NX-side integration, and the measured quality loop. Our job is the assistant your works actually needs — using Siemens components where they help.

Our documents are a mess — old scans, Word versions, three wikis. Precondition problem?

Messy sources are the normal starting point, not a blocker. The pipeline handles mixed formats and scanned PDFs, and revision logic ("rev F supersedes rev E") is part of the indexing. What genuinely helps is prioritising: the pilot starts with the two or three sources your engineers ask about most — typically works standards plus one archive — and expands from there. You do not need a document cleanup project first; the assistant often reveals what is worth cleaning up.

How does new knowledge get in — and old knowledge get corrected?

Connected sources sync automatically — a new works standard revision is answerable the day it lands in Teamcenter. Corrections flow through the feedback loop: engineers flag weak answers, and those flags drive retrieval tuning and, where the real cause is a contradictory document, point your standards owners to exactly what needs fixing. The assistant becomes, as a side effect, a quality audit of your documentation.

Herbert retires in spring. His answers don't have to.

Bring us the ten questions your seniors answer every week — we'll show you the assistant answering them from your own documents, with sources.

Talk to a Knowledge AI Engineer

Always here to support you – reach out to us with your questions, and we’ll guide you every step of the way.

Phone: +40 774 463 429
Our Location: Str. Ploiești 9, 400157
Cluj-Napoca, Romania
Email: info@inteliscience.net
LinkedIn: inteliScience on LinkedIn
YouTube: inteliScience on YouTube