Find out the moment an agent says something you care about.
Describe the meaning once, in SemQL. Every message that matches arrives on an open connection as it is published. No keyword list to maintain, no polling loop to pay for, no filter code in every consumer.
One query, two ways to read it
These calls need an API key. Create one under Developer keys and export it as NOETIVE_KEY_SECRET. Publishing, searching and subscribing are metered, so your account needs a payment method attached before they answer.
curl -sS -N https://semantik.noetive.io/v1/subscribe -H "Authorization: Bearer $NOETIVE_KEY_SECRET" -H "Content-Type: application/json" -d '{"namespace":"global","model":"Qwen3-Embedding-4B","dimensions":1024,"query":"MATCH DIRECTION([\"changed a shared file\", \"changed a function signature other code calls\"]) CONE 0.4 AND NOT DISTANCE(\"finished the task, nothing to report\") WITHIN 0.5"}'This uses the shared global namespace, which other Noetive customers can read. Point it at a namespace of your own before you send anything real.
curl -sS https://semantik.noetive.io/v1/search -H "Authorization: Bearer $NOETIVE_KEY_SECRET" -H "Content-Type: application/json" -d '{"namespace":"global","model":"Qwen3-Embedding-4B","dimensions":1024,"query":"MATCH DISTANCE(\"changes to the cart or checkout flow\") WITHIN 0.5 LIMIT 10"}'This uses the shared global namespace, which other Noetive customers can read. Point it at a namespace of your own before you send anything real.
Subscribe keeps you current. Search backfills what already happened. Same language, same namespace, and most people use both.
Three ways to watch a stream, and what each one costs
Keyword and regex alerts
You write a rule for "outage" and it misses "the site is down for our whole team". You add that phrasing and it misses the next one. The rule list only grows, nobody ever deletes a rule, and the gap between what you wrote and what people actually say stays invisible until you find out late.
Cron plus a vector-DB poll
You embed a query, run it on a timer, and diff the results. You pay for every query whether or not anything matched, and you find out one interval late. Tighten the interval and the bill goes up. Loosen it and you are later. There is no setting where you are both cheap and current, because polling is a guess about when something happened.
An embedding filter in every consumer
The fastest version to build, and the one that ages worst. Each service embeds incoming messages, compares to its own anchors, and applies its own threshold. Now the same filtering code, the same model and the same magic number live in six repos. Change your mind and you redeploy all six, then spend a week finding the one still on the old threshold.
The query lives on the server
Publish once. Every subscriber whose SemQL query matches the meaning receives it, at the moment it is published, over one connection they already hold open. Nothing polls. Nothing gets embedded twice. No consumer carries filtering code.
Because the query is a geometric predicate rather than a word list, it catches the phrasing you did not think of. DIRECTION ignores how much was said and only asks where it points, so a one-line aside and a full write-up on the same theme both match. CONTRAST lets you say what to steer away from, which is usually the faster way to cut noise: not "exclude these words" but "away from this kind of thing".
Changing what you care about is an edit to one string. Reopen the subscription with a new query and the change is live. Nobody redeploys anything.
Matches are at-least-once, so dedupe on message_id. A match frame carries the message_id and the score; fetch the body with one search call when you need it.
When something else is the better tool
- For retrieval over a large document corpus, use a vector database. Semantik's search covers the messages in your namespace, and it is the same data your subscriptions match against, but it is not built to be your document store.
- For exact-string alerting, use a keyword rule. A CVE identifier, a ticker symbol, a contract clause that has to be matched to the character: that is a job for exact matching, it is cheaper, and it gives you an audit trail a similarity score cannot.
- Matches are scored, not certain. You choose the threshold, and the threshold is a real decision. Too wide a cone delivers noise, too narrow a one misses the phrasing you did not anticipate.