Troubleshooting

The most common questions and what to do about them. Email support if your problem isn't here.

"Sign in with Microsoft" hangs or returns an error

Most common cause: your organisation requires admin consent for new applications. Knowledge requests only the minimum scopes (sign-in + basic profile), but corporate policy may still require an IT-admin tick.

What to do:

• Forward the consent URL (in the error page) to your IT / Entra admin. One click on their end and sign-in works for everyone in your tenant.
• If you get a different error, the URL bar usually contains an error_description query parameter — paste it into an email to support@olyteck.com and we'll tell you what to do.

Upload is stuck at "queued"

The background worker isn't running, or the queue is backlogged.

• Most of the time the worker starts within ~1 second of upload (we kick it from the upload endpoint). If you see queued for more than ~30 seconds, refresh the Jobs page — it'll catch up on the next cron tick (within a minute).
• If it's stuck for several minutes, contact support — we monitor the worker but occasionally something needs a kick.

Upload failed: "Extractor produced no text"

The file is probably a scanned PDF (image-only, no text layer) and OCR couldn't find any text. Two paths forward:

• Check OCR language settings. The default pack is English + French. If your document is in another language, ask support to enable the relevant Tesseract language pack — usually a 5-minute change.
• Check the scan quality. OCR struggles with low-resolution or skewed scans. If the file is poor quality, re-scanning at 300 DPI usually fixes it.

"Search returns nothing" — but I know the answer is in there

Five things to check, in order of likelihood:

1. Project scope. Is the search scoped to a project that doesn't contain the document? Clear the scope and try again.
2. Permissions. Are you a member of the project? Members only see projects they belong to. (Workspace admins see everything.)
3. Did the doc finish processing? Library → check the document's status. Documents only become searchable once status is completed.
4. Wording. Try synonyms. Semantic search handles paraphrase but isn't magic — "passive income" won't find a doc that's only ever called it "interest revenue."
5. Embedding failed. Rare. If a chunk's embedding errored at indexing time (e.g. Gemini rate limit), the fallback FULLTEXT pool should still find it on keyword match. If a tenant admin runs the diagnose page they'll see counts of missing embeddings.

The answer is wrong — the model invented something

This is what citations are for. Open the cited source: if the passage actually says what the model claimed, the answer is grounded but maybe wrong (the source is wrong). If the passage doesn't say what the model claimed, that's hallucination — three things to do:

• Re-ask with the question phrased more narrowly. "Do we encrypt backups?" beats "what's our security posture?"
• For an agent, turn grounding mode ON (the default for new agents). The unbreakable rule forces the model to abstain when the corpus doesn't cover the question.
• Try a different model. Premium-tier models (Claude Sonnet especially) abstain more reliably than cheap ones. See Pick the right model.

"402 — plan limit reached" on upload / search / agent run

You hit a quota for your current plan. The toast usually says exactly which one (documents, storage, generations, projects, members). Two paths:

• Make room — see deleting documents. Generations and storage reset / recompute on usage.
• Upgrade plan — account popover → Upgrade plan. Takes effect within seconds of the Stripe confirmation.

Office add-in: "401 — unauthorised" / "Agent not found"

• The key was revoked (Agents → key list → if there's no row, it's gone).
• The agent was deleted or disabled (Agents → check status).
• The token in the add-in's Settings tab is mis-pasted — re-mint and paste again.

Excel add-in: rows process slowly

Each row is a separate API call to a language model — there's a hard floor of 1-3 seconds per row (standard model) or 5-15 seconds (premium model). For a 200-row questionnaire that's 10-50 minutes total. Behaviour, not a bug.

If it's much slower than that:

• Check the per-key rate limit on the agent — if you set a low cap, the add-in respects it.
• Check the model — premium models are slower; switch to standard for bulk runs (see Pick the right model).
• Check Excel itself — very large workbooks (50,000+ rows) slow Excel down independently.

"I can't find the Word / Excel ribbon button"

The add-in was installed but Office hasn't picked it up yet. Two fixes:

• Restart Word / Excel.
• If still missing: Insert → My Add-ins → Manage My Add-ins. The add-in should be listed. If it's missing, re-run the steps in Install the add-ins.

"Removed a teammate but they still appear"

Their session is invalidated on the next request, but cached state in their browser doesn't clear immediately. They'll be signed out within 30 seconds and won't be able to sign back in. If you have an immediate concern (e.g. departure under bad terms), additionally rotate any agent keys they had access to (Agents → key list → revoke).

"Where do I download the data we put into Knowledge?"

Two paths:

• Individual documents: Library → open document → Download original. You get the exact bytes we received at upload.
• Bulk export of metadata + chunks: ask support@olyteck.com. We hand you a JSON archive within 5 business days. This is part of your GDPR data-portability rights — no extra fee, no contract change.

Still stuck?

Email support@olyteck.com with:

• What you were trying to do.
• What you saw instead.
• Screenshot if at all possible (drag & drop into the email).
• Document name (we don't read the contents — just need to look up the row).
• Approximate time when it happened (so we can correlate with logs).

Weekdays we typically respond within a few hours.