Commerce,in vectors.
Search. Cluster. Rank. Match.One embedding model.
Try the playgroundFind what they mean,
not what they typed.
Exact when it must be,
semantic when it can be.
One query, many intents.
Say what relevant means.
Ask in any language.
Find it in every one.
Sort any list by what
matters to the shopper.
How alike are these two?
One number.
Find the groups
nobody told you about.
See the shape
of your catalog.
Label anything. Bring
your own labels, or none.
Same product,
different seller.
More like this, and like that,
but not like that.
Same answers, a fraction
of the bytes.
toqan embed see
Images and text in one space. The placeholder model is text-only, so there is no demo yet.
toqan embed answer
Answers over specs, reviews and policy, with the retrieved passages shown first.
Cross-encoder reranking
The rank tab sorts with embeddings only. We do not claim reranker quality until we have measured it.
Playground
Connecting…Developers
One endpoint.
Twelve ways to use it.
Confirm you are human and get a key that starts with tmp_toqan_. No sign-up. It works for every endpoint for one hour. The code examples fill it in for you. Without a key you get 60 requests per minute.
| Texts per call | 128, up to 8,000 characters each |
| Body size (task endpoints) | 64 KB |
| Dimensions | 32 to 4,096 (default 1,024) |
| Inline corpus | 200 items |
| Model | Placeholder: Qwen3-Embedding-8B |
Pattern: embed with the API, cluster in your code
Questions
Good questions.
Clear answers.
Which model is running?
A placeholder: Qwen3-Embedding-8B, hosted on Nebius, used until toqan embed finishes training. Every response carries meta.placeholder: true and this page says so too. We show no toqan embed scores before training finishes.
Is everything on screen real model output?
Yes. Anything you type is embedded live. The sample catalogs are synthetic, and their vectors were computed once with the same model and stored. Numbers come from API calls or from math over those real vectors. Each tab shows where it struggles, with measured numbers.
How do I call the API from my own code?
In Developers, confirm you are human to get a temporary key (tmp_toqan_…). Send it as Authorization: Bearer <key>. It works for one hour, up to 1,000 requests per minute. Without a key, requests are limited to 60 per minute per IP. /api/v1/embeddings is OpenAI-compatible, so the OpenAI SDK works with a base_url override.
Why does cosine change when I add an instruction?
Queries can carry an instruction that tells the model what relevant means; documents never do. Scores shift with the instruction more than ranks do, so the playground shows rank and normalised bars first and the raw cosine as the number.
Can I use it on my own data?
Most tabs accept pasted lines or inline items (up to 200). For larger sets, embed with the API and run the grouping in your own code; the Developers section shows the pattern.