What it will not do, what happens instead, and where each one is a limit of document search in general rather than of this product.
- It won't tell you which revision is current
- If both revisions are in the workspace, you can be handed the superseded one — and the citation will look right, because that document is real. Revision tracking is on the roadmap. Until it ships: every answer names the document and page, so you are checking a named file rather than trusting a paragraph, and keeping one current revision in the workspace — superseded ones in a separate folder — makes the cited one the current one.
- Also true ofNothing consumer-grade does this either. ChatGPT has no way of knowing which of two files you uploaded supersedes the other, and your shared drive will hand you a superseded WPS without comment. What you get here is the document and page behind every answer, so a superseded file is visible instead of silent.
- It won't promise you a complete list
- Ask for every job that used heat 38291 and what comes back is the most relevant passages, not a guaranteed complete set — treat it as a starting point, not a final count. It says so rather than handing you a short list that reads like the whole one. Ask about a single job, certificate or product and it answers from the passages that match, with the page.
- Also true ofEvery search you already use works this way: ChatGPT answers from what it found, and so does the search box on your shared drive. Neither one warns you that the list may be short.
- It hasn't been tested on your kind of documents yet
- Our published tests cover legal contracts, FDA drug labels and SEC filings, not fabrication quality records. Try it on your own files during the trial and check the cited pages — that is the only test that counts for you.
- Also true ofWe have not found a chatbot that publishes how it performs on documents like yours. Ours are public, including the hard suites where we score badly — which is why we can tell you what has not been tested yet.
- Scanned certificates are the weak spot
- Scanned certificates are read by OCR, and a table in a scan can come back imperfect. Every answer names the page it came from, so confirming a value against the original takes seconds — the same check you would do before quoting it to an auditor.
- Also true ofOCR misreads are universal, ChatGPT included. The difference is what happens next: a misread number here arrives with the page it was read from, so it can be caught.
- It doesn't sign anything off
- It puts the procedure, the certificate and the qualified range in front of you in seconds, with the page, so the call is made on the document rather than from memory. The responsible welding coordinator still makes it — and now has the evidence to show for it.
- Also true ofSigning off is a competent person's job, not software's. What this has to do is show you the document behind the answer when the notified body asks you to justify the call.