The knowledge is real. It's just held by one person, or a filing cabinet.
Most manufacturers that have been running for a couple of decades sit on an enormous amount of accurate, hard-won knowledge — which revision of a drawing is the one actually running on the floor, why a process sheet has a step nobody wrote a reason for, which supplier's material occasionally runs out of tolerance, what a customer's non-standard part number from an old ERP system actually maps to. None of that is missing. It exists in a wiki page nobody updated after 2015, a shared drive with folders named after people who left years ago, a filing cabinet of paper travelers, an old PLM system almost nobody still has a login for, and — more than any of the systems — in the heads of a handful of engineers, quality leads and shift supervisors who have been at the plant long enough to just know.
That last part is the real exposure. A planner who has run the same line for thirty years knows without looking that a particular part needs a second inspection pass, or that a certain supplier's castings need extra scrutiny in the summer months. None of that is written down anywhere, because it never needed to be — the person was always there to ask. When that person retires, the knowledge does not get handed over in an exit interview. It just stops being available, and the plant finds out it's gone the first time a new hire makes the mistake the retired engineer would have caught without thinking.
A knowledge system does not fix a documentation problem that was never solved. What it does is make everything that already exists — the good drawings, the accurate process sheets, the supplier certificates, the ten years of quality records — findable by the people who need it, in the moment they need it, instead of only by the one person who happens to remember where it lives. Where knowledge exists only in someone's head, the same system gives you a structured way to capture it before that person leaves, rather than losing it the day they clean out a desk. It's one specific, deep version of the broader case for AI in manufacturing operations — finding what your plant already knows, faster than a person can.
What the system does on a normal day
The shape of the work stays consistent even though the source material never does.
- Engineering drawing and revision retrieval. Someone asks for the current drawing on a part — by part number, by a customer's own number, by a rough description of what it is — and gets back the correct current revision, not the first file that matches, with every superseded version clearly marked as superseded rather than quietly returned as if it were live.
- Process and work-instruction lookup. A shift supervisor or a new operator asks how a step is actually done, and gets the current work instruction for that operation, on that line, for that part — not a generic version, and not whatever printout happens to be taped to the machine.
- Supplier spec and material certificate retrieval. A quality engineer or a buyer needs the certificate of conformance, the material spec, or the test report tied to a specific lot or purchase order, and gets it pulled from wherever it actually lives — a supplier portal export, an email attachment, a scanned PDF in a shared drive.
- Quality and NCR history — "has this happened before?" A new non-conformance comes up, and instead of relying on someone remembering a similar case from years back, the system searches prior non-conformance reports and corrective actions for the same part, process or failure mode and surfaces what it finds.
- Maintenance and equipment manuals. A technician troubleshooting a fault on an older machine gets the relevant section of the manual, the maintenance history, and any internal notes on that specific piece of equipment, instead of paging through a PDF that was scanned once and never indexed.
- Capturing what senior staff know before they leave. For the knowledge that was never written down at all, we build a structured capture process — recorded conversations, guided write-ups against a template — that turns what's in someone's head into a searchable record before it walks out the door.
Three retrievals that used to depend on one person
The clearest way to see what this looks like is to walk through it.
The correct revision of a drawing nobody has quoted since the old ERP
A customer calls asking for a repeat order on a part they last bought seven years ago, under a part number that predates the plant's current ERP. An estimator asks the system for the current drawing. It searches the file servers, the PDM exports and the scanned archive together, finds three drawings that match the part description, and returns the most recent revision — flagging the two earlier ones explicitly as superseded, with the revision letter and date of each version shown, not hidden. The estimator opens the current drawing, confirms it against the customer's reference sample, and quotes the job the same day instead of losing it to the afternoon it used to take to track down which folder held the right file.
A supplier certificate an auditor wants for a lot shipped last year
An auditor asks for the certificate of conformance tied to a specific lot of raw material used in a customer's order from the previous year. The quality lead asks the system, which pulls the purchase order, matches it to the supplier's certificate on file — an email attachment from over a year ago — and hands back the document along with the lot number and receipt date it's tied to. If the system cannot find a certificate for that exact lot, it says so plainly rather than returning the closest match as if it were the right one, because a wrong certificate in an audit file is worse than a flagged gap.
"Has this failure happened before?" — a quality engineer checking NCR history
A new defect turns up on a part midway through a run. Before writing it up as a new issue, the quality engineer asks the system whether this failure mode has shown up before on this part or process. It searches the non-conformance history and surfaces two prior cases — one from a few years back with a documented root cause and corrective action, one more recent that was closed without a clear cause identified. The engineer now has real history to work from instead of starting the investigation from nothing, and can decide whether the earlier corrective action actually held or needs revisiting.
What this doesn't fix — and where it gets hard
Being honest about the hard parts here matters more than almost anywhere else, because a knowledge system that gets manufacturing documents wrong is genuinely dangerous, not just unhelpful.
Scanned drawings and old paper travelers need OCR to become searchable, and the scan quality on a document that's been in a filing cabinet for twenty years is often poor — faded lines, handwritten revision notes, a stamp obscuring a dimension. OCR on material like that is not perfect, and we tell you upfront where confidence is low rather than presenting a shaky read as a clean one.
Revision control is the harder problem, and it has to be respected structurally, not left to chance. A system that confidently hands someone a superseded drawing is worse than a system that finds nothing at all, because a person who gets no answer knows to keep looking, while a person handed the wrong revision has no reason to doubt it. We build revision handling in from the start — tracking which version is current in the source system of record, surfacing supersession explicitly, and refusing to guess when two versions look equally plausible and the metadata doesn't clearly say which one is live. Every answer the system gives cites the specific document and revision it came from, so a person can verify it against the source rather than taking the system's word for it.
And knowledge capture from senior staff is a project in its own right, not a feature that runs itself. It takes deliberate time with the people who have it, before they leave, and it works best started well before a retirement date is set rather than in the last few weeks.
Connecting to your PLM, ERP and shared drives
This is the same discipline behind Calfy's knowledge systems work generally, applied to the specific mess a plant actually has.
Connection. We index what you actually run — file servers, PDM or PLM exports, an ERP's document attachments, scanned archives, shared drives with a decade of undocumented folder structure. You do not need to migrate everything into a new system before this works; the system searches across what exists today.
Permissions. Access mirrors what your source systems already enforce — a supplier's certificate is visible to the people who could already see it in your quality system, not opened up to everyone with a login. Every retrieval is logged against the identity that made it.
Where OCR and document review matter most. For scanned drawings, supplier certificates and paper NCRs, the underlying AI document review work is what makes the archive searchable in the first place — extracting text, part numbers, revisions and dates from documents that were never digitized as data.
Retrieval, not invention. The technique underneath this — search your actual documents first, then answer from what was found, with the source attached — is the same retrieval-augmented generation approach behind any properly grounded system, and the knowledge base search use case covers the general pattern in more depth.
Where it sits next to an agent. A knowledge system answers questions; it does not act on them. Where a plant also wants something that reads an RFQ or a change notice and does the next step in the ERP, that is the territory covered by AI agents for manufacturing — often built on top of the same indexed document base.