The right revision, the exact clause, and the reason behind the decision — with the source open in front of you.
One private knowledge base across your specifications, drawings, calculations, change orders, test reports, RFIs, e-mail and ERP. Every answer cites the passage it came from, says whether that source is current, and tells you when two documents disagree. Your project data stays under your control, and Olyra operates it as a managed service.
What is Olyra Knowledge Engine for engineering firms? It is a private, permission-aware knowledge base built for work where exact details matter. It reads your project records — specifications, standards you hold, drawings and their revisions, calculations, change orders, test reports, RFIs, e-mail and ERP records, including scanned and legacy paper — and lets engineers and AI agents ask questions in plain language. Every answer cites the exact source passage and opens the document. Each source shows whether it is verified, stale or superseded. Conflicting values between documents are flagged, and when the evidence is missing the engine says so instead of guessing. You can also search the records as they stood on any past date. Your data and keys stay with you; Olyra installs it and operates it as a managed service that keeps it current, secure and supported.
In engineering, an almost-right answer is a wrong answer
The knowledge exists. It is in a project folder nobody has opened in years, in a mail thread with the client, in the ERP, or in the head of the one person who was there. These are the sentences we hear from engineering leads.
“We quoted from Rev B. The client was on Rev D.”
Nothing in the folder said the note had been superseded. The old value was copied into a proposal with full confidence, and the difference surfaced after the order.
“Why did we choose that material?”
The reason is in a mail thread from three years ago, or in a senior engineer’s memory. The drawing records what was decided, not why.
“Ask Paulo, he did the last one.”
A few experienced people carry the exceptions, the client history and the lessons learned. When they retire or are simply on site, the answers go with them.
“Have we done something like this before?”
Junior engineers cannot find precedent, so they start from zero or interrupt someone. Similar jobs sit in closed project folders with names only the original team understands.
“It is somewhere between the file server, e-mail and the ERP.”
Requirements arrive by e-mail, drawings live on a share, the bill of materials is in the ERP, and the clarification was agreed in a chat. No single search covers all four.
“The chat tool gave us a tolerance that is in no document.”
A general assistant produces a plausible, well-written answer with no source behind it. For exact-detail work that is worse than no answer, because somebody may use it.
Clients add a further demand: show where each requirement came from and which revision it was checked against. Assembling that trail by hand is slow, and it depends on the same few people.
What changes when the records answer for themselves
The questions stay the same. What changes is that the answer comes with its evidence, and with a clear signal when the evidence is weak.
| The question | Today | With Knowledge Engine |
|---|---|---|
| “What is the current design pressure for this skid?” | Open three folders, compare file dates, ask a colleague to confirm | An answer citing the passage in the current datasheet, with the superseded revision marked as superseded |
| “What did the client specification say when we issued the quote?” | Reconstruct from mail attachments and memory | Search the knowledge base as of the quote date and read what was on record that day |
| “Why was the weld procedure changed on this job?” | Find the person who remembers | The change order, the mail thread behind it and any recorded precedent, each cited |
| “Do the datasheet and the drawing note agree?” | Discovered at inspection, or by the client | A conflict flag showing both values and both sources |
| “Is this proposal text consistent with our records?” | A senior engineer reads it line by line | A grounding check marks each sentence supported, contradicted, out of date or unsupported, before it is sent |
The examples on this page are illustrative. They describe the kinds of questions the engine is built for, not the results of a named client. The full product is described on the Knowledge Engine page.
Exactness: how the engine handles it
A general model is good at fluent language. Engineering needs the specific value, from the specific revision, and an open admission when it is not there. Six mechanisms do that work.
Citation to the passage
Not “see project folder”
- Each answer cites document, version and exact location
- One click opens the original with the passage highlighted
- Answers are built only from passages the asker may see
Status on every source
Verified, stale, superseded or disputed
- Each document carries a verification and freshness state
- An answer that leans on an old document says so, once and clearly
- Your data steward can mark a document superseded or retired, in bulk where needed
Conflicts are shown
Two values, two sources
- Disagreements between sources on numbers, dates and revision identifiers are flagged
- Both values are displayed with their citations
- The engineer decides which governs; the engine does not pick silently
It abstains
“Insufficient evidence” is a valid answer
- When the records do not support an answer, the engine says so
- A “who owns this?” hint points to the owning role or your data steward
- A verified-only mode limits client-facing agents to verified, in-date, internal sources
As-of search
What was true on a given date
- The whole knowledge base can be viewed as it stood on any past date
- Useful when a question turns on what the specification said at quotation or at design freeze
- The same question can be asked for today and for that date, and compared
Grounding check on drafts
Before the proposal or reply leaves
- Checks a draft sentence by sentence against your records
- A wrong number, date or revision is shown with both values
- Runs on the box with no AI model involved, and never edits or blocks on its own
What this does not do: Knowledge Engine checks only what has been ingested. It does not perform engineering calculations, it does not replace engineering judgment or professional sign-off, and it does not make a drawing or document authoritative on its own. Your document control procedure and your responsible engineer still decide what governs.
Chosen for engineering work
Each capability below answers one of the problems above.
| The pain | What the engine does about it |
|---|---|
| The wrong revision gets quoted | Freshness and verification status on every hit; superseded documents are labelled; revision conflicts between sources are flagged |
| The “why” is lost | Mail and chat threads are searchable next to the documents they explain. Precedents record the task, the final text and the reason, and curated answers approved by a steward take precedence for the questions that matter most |
| Senior engineers are retiring | Their projects, correspondence and corrections remain findable and cited, which reduces key-person risk and supports a cleaner handover |
| Juniors cannot find precedent | Search by meaning and by exact keyword together, filtered by client, project, document type, date or person; a map of clients, suppliers, people and products links related work |
| Knowledge is scattered | File servers, Microsoft 365 or Google Workspace, mail, chats and ERP records in one knowledge base. ERP records connect read-only: ready-made connectors for TOTVS Protheus, SAP Business One, Senior and Microsoft Dynamics 365 Business Central, and connectors for other ERPs such as Oracle NetSuite, SAP S/4HANA, Epicor, Infor and Sage built and added for each customer. Other systems such as a PLM can come in through an ODBC driver, an export, a webhook or the connector SDK |
| Old projects are on paper or scans | Built-in OCR reads scanned PDFs and images in the languages your documents use, so the text on scanned reports, datasheets and drawing sheets becomes searchable |
| Client work must stay separated | Permissions are enforced in the database itself, by person and by the categories you define, such as client and project. Existing file-share and SharePoint restrictions can be carried over |
| Clients ask for traceability | Citations to document, version and location; a receipt for what each question was shown; a tamper-evident audit trail of who asked what |
| AI agents need exact details mid-task | Agents connect through the open MCP standard under their user’s permissions, receive cited and status-checked context, and are told to stop when the evidence is missing |
For agents, speed and exactness arrive together
An agent drafting a quotation, answering an RFI or preparing a submittal list needs to look things up many times inside one task. If each lookup takes a minute, the agent is either slow or it stops checking.
Knowledge Engine prepares your data in advance into a database in several forms — full text, meaning, a graph of people and things, and structured records — so answers come from what is already built, fast enough to use in the middle of a task. Lookups can return in as little as a couple of milliseconds. Results vary with hardware, data volume and query type.
- One bounded briefing per client or task: the relevant records its user may see, approved answers, verified passages, precedents and warnings, with citations and a receipt
- Outside content, such as a supplier e-mail, is labelled untrusted so a stranger’s text cannot steer the agent
- Corrections your engineers make to an agent’s work become precedents the next agent can see, once a steward approves them
- Every agent call is logged in the audit trail
Any assistant that speaks MCP can connect, including Claude and Copilot-style tools. Olyra’s pre-built AI agents use the same interface.
A general chat-with-documents tool versus Knowledge Engine, for exact-detail work
Uploading a few files to a chat assistant is a sensible way to summarize a report. This is how the two compare when the task is the exact value from the governing revision.
| For exact-detail work | General chat-with-documents tool | Knowledge Engine |
|---|---|---|
| Where the answer comes from | The files uploaded for that conversation, plus the model’s general knowledge | Your ingested records only, cited to the passage |
| Which revision | Whichever file was uploaded; usually no signal that a later one exists | Status on each source, and a flag when revisions disagree |
| When the detail is not in the files | Often a fluent answer anyway | Abstains and says the evidence is insufficient |
| What was true at an earlier date | Not available unless you find and upload the old files | As-of search across the whole knowledge base |
| Client separation | Depends on what each person uploads and where | Permissions enforced in the database, per person |
| ERP records and scans | Typically outside its reach | ERP records beside documents; OCR on scans in your document languages |
| Record of use | Varies by tool | Tamper-evident audit trail and a receipt per question |
| Where it shines | Quick summaries and drafting on a handful of files | Daily exact-detail work across years of projects, for people and agents |
The other options an engineering firm usually considers
Each of these is the right answer for some firms. Many run alongside Knowledge Engine rather than instead of it.
| The alternative | Where it is strong | Where it tends to fall short for this job | How Knowledge Engine relates |
|---|---|---|---|
| Document control or PLM system | The authoritative record of released revisions and approvals | Holds the controlled documents, but not the mail, chats, RFIs and reasoning around them; search is usually by number and metadata | Does not replace it. Reads alongside it and connects the surrounding context, with citations back to the source |
| File-server or SharePoint search | Already there; no project needed | Keyword search inside one system; nothing says which version is current, and the ERP and mail stay separate | One cited answer across sources, with status on each |
| Copilot bundled with the office suite | Inside the tools people already use | Centered on that suite’s content and on today’s permissions, oversharing included | Adds file servers, scans and ERP records; Copilot-style tools can call it through MCP |
| A wiki or lessons-learned register | Good when someone maintains it | Depends on engineers writing things up after the job, which is when nobody has time | Works from the records the project already produced; curated answers and precedents add the written-up layer where it is worth it |
| Build it in-house | Full control; a capable team can show a demo quickly | It sees only the sources it was wired to, so it cannot tell whether the document it found is the current, approved or complete one, and it answers confidently from a superseded revision. Permissions, sign-in, audit, OCR, freshness, as-of search and upkeep are each a project of their own | Connects all your sources and records how current and how trusted each one is, so an answer comes with evidence of whether it is right. Those pieces ship together, and Olyra installs, updates and operates the system |
Cited answers to your own questions on day 3
We do not start with everything. We start with the one source where your engineers lose the most time.
Install and connect one source
Sign-in is federated with your identity provider. One high-value source is connected first: for many firms, the project file share for an active client or product line.
Your questions, answered
Ten real questions from your engineers, asked and answered in front of you, with citations that open the source.
Categories and a first assistant
Client, project, discipline and document-type categories are in place, more sources are connected, and a first assistant does one narrow job end to end.
The first evaluation report
A golden set of your own questions, measured results, and a data steward trained to run it.
The first ingestion is real work: your records are read, extracted, classified and built into a fast database in several formats, which takes time and AI processing. For that reason we do not set up a free trial on your own data. We can show you a demonstration box running on sample data, and the engagement itself is arranged so that you see cited answers to your own questions on day 3, before the remaining sources are connected.
Client drawings and specifications stay under your control
Much of what an engineering firm holds belongs to its clients or is covered by a confidentiality agreement. The design starts from that fact.
Your data, your keys
The data, every export and the recovery keys are yours. It all runs on a box in your building or a server in your own cloud account, not in a shared cloud service.
A managed service that keeps no copy
Olyra installs, updates, patches, monitors and supports the system, so it stays current and secure without your engineers maintaining it. Nobody has standing access: a support session is requested with a reason and a duration, approved by you, and recorded.
Models at addresses you control
AI models are reached on your own GPU, on a cloud service under your own account, or on the box. Olyra does not host AI inference, and your data is not used to train anyone’s models.
Residency rules
Rules decide what may leave the box, by where content came from or what it is — for example, everything belonging to one client. A simulator shows what a rule would allow before you switch it on.
Separation by client and project
Permissions are enforced inside the database, so no screen, integration or AI assistant can go around them.
Always exportable
Your data can be exported at any time, and a factory release exports it in full if you ever decide to stop.
More detail is in the Trust Centre and on the main Knowledge Engine page.
Nothing on this page is legal, tax or financial advice. Whether these controls meet a particular contract, standard or regulation is a question for your own counsel and your client; Knowledge Engine supplies records and evidence, not a compliance outcome.
One office or several
Most firms begin with a single box. From there you can add a second box as an automatic failover pair, build a cluster at one site for heavy load, place a replica at each office, and keep a disaster-recovery standby at another site. With several offices, one site is the master: a replica sends its local documents to the master, the master ingests them, and they return to the local site at the next replication. Nothing is reinstalled as you grow, and we size each stage with you.
Who it is for, and who it is not for
A good fit
- Engineering services, design-build, EPC and consulting engineering firms with years of project records
- Engineering departments of engineered-to-order manufacturers, where each order carries its own specification and change history
- Industrial, mechanical, electrical and civil design teams whose clients ask for traceability
- Firms facing the retirement of senior engineers, or onboarding many junior ones
- Firms that want AI agents to help with quotations, RFIs or reports and need those agents to cite exact sources
Probably not a fit
A small team whose records fit in a few hundred well-organized documents will be well served by a lighter tool, and we will say so. It is also not a calculation package, a CAD or model viewer, or a replacement for your document control system: it reads the text of your records and answers from them. And it needs a data steward on your side for a few hours a week in the first two months.
Questions engineering leads ask
Can it give a wrong answer?
It can only answer from what has been ingested, and a source can itself be wrong. What it does is remove the common causes of error: every answer cites its passage so you can check it, superseded and stale sources are labelled, conflicts between documents are shown, and it abstains when the evidence is missing. It does not replace engineering judgment or professional sign-off.
Does it read drawings?
It reads the text of documents, including scanned PDFs and images through OCR in the languages your documents use, so title blocks, notes, revision tables and scanned reports become searchable. It does not interpret geometry, check a design or perform calculations, and it does not make a drawing authoritative on its own.
We already have a document control or PLM system. Why add this?
Keep it. That system is the record of what was released. Knowledge Engine connects the released documents with the mail, RFIs, change orders, test reports and ERP records around them, and answers questions across all of it with citations back to each source.
How does it know which revision is current?
From the records and from your steward. Each document carries a verification and freshness state; a steward can mark documents verified, superseded or retired; and the engine flags sources that disagree on revision identifiers, numbers or dates. Where your own procedure defines what governs, that procedure still decides.
How is confidential client work kept apart?
You define categories such as client and project, and permissions are enforced in the database itself for every person and every AI agent. Existing file-share and SharePoint restrictions can be carried over, and an access review shows who can see what, and why.
Does our data leave the building?
Only where a rule you set allows it. With the GPU configuration, models run on the box. Otherwise models are reached at a cloud service under your own account, in the region your policy allows. Olyra holds no copy of your data and does not host AI inference.
Can we try it on our own drawings first?
We can show you a demonstration box on sample data. We do not offer a free trial on your own data, because the first ingestion is real work: reading, extracting, classifying and building the database takes time and AI processing. Instead, the engagement starts with one high-value source and shows cited answers to your own questions on day 3.
How fast is it?
Fast enough to use in the middle of a task, because your data is prepared in advance. Lookups can return in as little as a couple of milliseconds. Results vary with hardware, data volume and query type.
What do you need from us?
A sponsor, a data steward for a few hours a week during the first two months, read access to the first source, somewhere to put the machine or a cloud account, and a decision on data residency.
Bring ten questions your engineers ask every week
In a 30-minute briefing we look at where your project knowledge lives, which source would show value first, and what your ten questions would look like answered with citations. If a lighter tool would serve you better, we will tell you.
Book an engineering knowledge briefing
Prefer to read first? The full product description is on the Knowledge Engine page.
Find out what AI can actually do for your operations.
Start with an AI Operations Audit — a fixed-scope diagnostic that maps your workflows, scores your AI readiness, and hands you a prioritized 120-day roadmap. If we don't find real opportunities, you'll know that too.
Request an AI Operations Audit → Prefer to talk first? Book a 30-minute scoping call.