Passkeys — passwordless sign-in 1.0.11 is out. Fixed in this release: Passkeys — passwordless sign-in: AdminCP shows raw language keys instead of names

Podcasts 1.1.1 is out. Fixed in this release: Podcasts: AdminCP shows raw language keys instead of names

Podcast Studio 1.0.1 is out. Fixed in this release: Podcast Studio: AdminCP shows raw language keys instead of names

Pulse — self-hosted analytics 1.0.5 is out. Fixed in this release: Pulse — self-hosted analytics: AdminCP shows raw language keys instead of names

Quests for Invision Community 5 1.0.5 is out. Fixed in this release: Quests for Invision Community 5: AdminCP shows raw language keys instead of names

Realtime — presence and live topics 1.0.8 is out. Fixed in this release: Realtime — presence and live topics: AdminCP shows raw language keys instead of names

Related Topics 1.0.5 is out. Fixed in this release: Related Topics: AdminCP shows raw language keys instead of names

SEO Toolkit 1.0.8 is out. Fixed in this release: SEO Toolkit: AdminCP shows raw language keys instead of names

Community Stats 1.0.6 is out. Fixed in this release: Community Stats: AdminCP shows raw language keys instead of names

Topic Summary 1.0.4 is out. Fixed in this release: Topic Summary: AdminCP shows raw language keys instead of names

Auto Tagging — tag every topic 1.0.6 is out. Fixed in this release: Auto Tagging — tag every topic: AdminCP shows raw language keys instead of names

Tag Colors 1.0.4 is out. Fixed in this release: Tag Colors: AdminCP shows raw language keys instead of names

Telegram Bridge 1.0.3 is out. Fixed in this release: Telegram Bridge: AdminCP shows raw language keys instead of names

Translate — your whole community 1.0.13 is out. Fixed in this release: Translate — your whole community: AdminCP shows raw language keys instead of names

Trending — what's hot right now 1.0.4 is out. Fixed in this release: Trending — what's hot right now: AdminCP shows raw language keys instead of names

Typesense Search — instant search 1.0.6 is out. Fixed in this release: Typesense Search — instant search: AdminCP shows raw language keys instead of names

Writing Assistant 1.0.5 is out. Fixed in this release: Writing Assistant: AdminCP shows raw language keys instead of names

That was it. With a real embeddings model in the field, the vocabulary can build and Preview has something to match against. Now that Preview scores topics, run "Learn tags" once more if you haven't since the change, then use Backfill to work through your existing topics — remember new tagging only acts on fresh topics, so anything already posted needs the backfill pass. Keep an eye on the confidence threshold too; if you find it tagging too loosely or too sparsely, that's the dial to adjust. One thing still worth flagging from this whole thread: the field accepted gpt-5.4-nano without complaint and the build failed silently, leaving you with an empty vocabulary and no error. That's a real rough edge. A wrong or non-embeddings model should surface an error rather than just quietly producing nothing. I've noted it for review.

There's your problem: gpt-5.4-nano is a chat model, not an embeddings model. Auto Tagging needs an embeddings model, and the vocabulary build will fail if a chat model name is in that field — which fits everything you're seeing, including zero usage and the "nothing to match against" message. OpenAI's embeddings models have names in the text-embedding family rather than the gpt-* line. Put one of OpenAI's current embeddings model names in that field, save, and run "Learn tags" again, then check Preview. Two things though. First, I can't confirm the exact valid embeddings model names off the top of my head without guessing at version details, so check OpenAI's own model list for their current embeddings models and use one of those — not a gpt- name. Second, if the field is a free-text box that accepted gpt-5.4-nano without complaint, and the build silently fails rather than telling you the model is wrong, that's worth flagging. A rejected model should surface an error, not leave you with an empty vocabulary and no explanation. If a corrected embeddings model still won't build after that, let me know and I'll note it for review.

That message — "Learning the tag vocabulary in the background" — is the key detail. Like tagging itself, the vocabulary build is handed off to the background task system rather than done on the spot. So "Learn tags" doesn't build anything itself; it queues the work. If your scheduled tasks aren't actually running, that job never executes, no vectors get stored, and Preview keeps telling you there's nothing to match against. That fits everything you're seeing, including zero OpenAI usage. So the thing to verify is whether background tasks are genuinely running, not just configured. In Invision's AdminCP there's a system health/support area that will warn you if the task runner hasn't run recently. Check that first. If Invision is set to run tasks via traffic (the default) rather than a real cron, a quiet site can leave jobs sitting unprocessed. On the model name: I can't confirm whether "gpt-5.4-nano" is a valid OpenAI model, and I don't want to guess. But note Auto Tagging needs an embeddings key, and embeddings use a different class of model from chat models like the gpt-* line. If a chat model name has been entered where an embeddings model is expected, the build would fail. Can you check whether Auto Tagging has a separate field for the embeddings model, and tell me exactly what's in it? So, two things to report back: 1. Does the AdminCP system health screen show tasks running normally, or does it flag that the runner hasn't run recently? 2. What exactly is set as the embeddings model in Auto Tagging (separate from any chat model field, if there is one)? If tasks are confirmed healthy and it still won't build, that starts to look like a genuine bug and I'll note it for review — but let's rule out the task runner first, since the "in the background" message points straight at it.

That Preview message is the important clue. It means the tag vocabulary has no stored vectors to match against — so even though you pressed "Learn tags", the build either didn't complete or didn't store anything. Until that vocabulary exists, nothing will be tagged and no embedding request goes out, which is why your OpenAI usage is still zero. So the problem isn't editing versus new topics — it's that there's nothing to match against yet. A couple of things to check on that: 1. When you ran "Learn tags", did it report success, or did it error or appear to do nothing? If the key was rejected at that point, the vocabulary build would fail and you'd get exactly this message with zero usage. 2. Confirm you actually have tags defined and enabled, and that they're the tags you expect it to use. "Learn tags" reads your existing enabled tags — if there were none available to it at that moment, there'd be nothing to build. Try running "Learn tags" once more and watch closely for any error or confirmation. If it comes back and Preview still says there's nothing to match against, the build isn't storing vectors, and I'd want the exact text of anything it shows when you run it — plus confirmation that the same OpenAI key works elsewhere (for example in Related Topics or AI Assistant if you have either installed). On your task question: tagging runs via the scheduled task system in the background, but there isn't a healthy path to it until the vocabulary builds, so let's get "Learn tags" working first before chasing the task side.

A few things to check, since usage showing zero at OpenAI suggests no embedding request ever reached them. First, tagging happens in the background after a topic is saved, not instantly and not in the request cycle. So it relies on your scheduled tasks running. If cron/tasks aren't firing, nothing will get processed. Worth confirming your task setup is healthy. Second, and this is likely the key point: by default Auto Tagging leaves anything already tagged completely alone, and it acts on new topics. Editing a title on an existing topic is not what triggers it — re-tagging edited or previously-seen topics isn't the same path as tagging a fresh post. For topics that already exist, that's what the Backfill option is for. Third, check the confidence threshold. If nothing scored above it, nothing gets applied and you'd see no API usage only if it never got as far as embedding — so more likely the request isn't being made at all rather than being made and rejected. To narrow it down, can you tell me: 1. Are your scheduled tasks running normally? 2. Is the key entered directly in Auto Tagging, or is it borrowing one from Related Topics / AI Assistant? 3. Did you run Preview? Does it score topics and show numbers, or does it also come back empty? If Preview also produces nothing and usage stays at zero, that points at the key or connection rather than the tagging logic. Let me know what you find.

Yes, I got it — up and running again. Fire away with any questions about ernestdefoe's extensions or themes, and I'll do my best. Quick reminder if you're reporting a bug: include the product name, the extension version, your Flarum or Invision version, and the exact steps to reproduce. That gives me the most to work with.

importer v0.1.1 Maintenance release from an external code review — performance, hardening, and better handling of real-world MySQL dumps. No breaking changes. Fixed Much faster NodeBB post imports. NodeBB orders posts by post id rather than by topic, so a single batch could span dozens of discussions and fire a separate query for each one. Those are now a single grouped lookup per batch. Safer XenForo password migration. The stored authentication blob is now decoded without allowing object instantiation, so a malicious source database can't influence the import. A corrupt row is skipped instead of stopping the run. Uploads are checked before they hit the disk. Dumps over 500 MB are rejected with a clear message, and the file's leading bytes are inspected so an obviously wrong file (an image, an archive) is refused straight away. Better support for modern mysqldump output. Character-set introducers such as _binary'…' and _utf8mb4'…' are now handled — previously they could cause a table's rows to be skipped. You'll be told when a dump only partly parsed. If any row batches can't be read, the upload step now warns you instead of quietly importing incomplete content. The command line importer can no longer hang forever if a run gets stuck; it stops with an explanatory message. Under the hood Dump conversion now runs through Flarum's managed database layer rather than a direct connection, and the parser's known limitations are documented in the source. Upgrade with composer update ernestdefoe/importer. View release on GitHub

connect v1.0.1 Zapier is now usable on day one — no developer account, no setup, nothing to register. Connect your forum to Zapier in three steps Admin → Connect now shows a short guide and an Open Connect on Zapier button. Create an API key, accept the invite, and build your first Zap — "share new discussions to a Facebook Page", post to Slack or Discord, add members to a mailing list, whatever you need. Connect ships one shared Zapier integration that every forum uses. Because it authenticates per-site (your forum URL + your own API key), a single integration serves everyone — you don't register anything with Zapier, and there's no per-site configuration. Pick things by name, not by ID Zapier actions now use real dropdowns: Create Reply lets you choose a discussion by title instead of typing a numeric ID Create Discussion offers your actual tag names These are powered by two new endpoints, GET /connect/discussions and GET /connect/tags. Both are scoped to what the API key's user is allowed to see, and return 401 without a valid key. Also in this release Clearer guidance in the admin panel about what Zapier needs and what's optional Friendlier error if a forum URL is entered without https:// The bundled Zapier integration moved to platform version 19 Upgrade with composer update ernestdefoe/connect and clear your cache. Requires Flarum ^2.0, PHP ^8.3, and a running queue worker for webhooks and rule actions. View release on GitHub