Embeddings

Consult the limit assigned to your key and the selected model’s batch-size limits. Control request concurrency and use backoff when the service returns a rate-limit response.

Training methods differ by generation. The linked papers explain the data mixtures, task adapters, distillation, and multimodal alignment used for each model.

v5-omni accepts text, images, audio, video, and PDF input. Other models support a narrower set; the playground only offers input types published for the selected model.

Support varies by model. Use the model card for the published language coverage and evaluate representative queries from your own corpus.

The catalog displays each model’s context window. Hosted API limits can be lower and may also depend on input modality and batch size.

The active API schema and service limits determine the batch size. Split large collections or use the asynchronous batch API when available.

Select a model that supports the modality, choose the input type, and provide the supported URL or data representation. The generated request shows the exact structure.

Evaluate retrieval on your queries and documents. Published benchmarks provide context, but latency, storage, language, and domain fit also affect the decision.

Re-embed the documents and rebuild the index when the embedding space changes. Vectors from different models are not generally comparable.

Image, audio, and video tokenization depends on the model and preprocessing. Review the response usage fields and the current provider pricing.

Use v5-omni for several media types, v4 for its multimodal and multi-vector capabilities, or a CLIP model for text–image tasks. Compare performance on your data.

Classifier training is available through the classification workflow. Updating the underlying embedding weights is a separate model-training task.

Check the model card for available weights, supported runtimes, and licensing. A hosted API subscription does not automatically grant a commercial weight license.

It selects a representation suited to a workflow. For retrieval, encode the query and the document with their corresponding retrieval tasks.

Instead of one vector per document, token-level vectors preserve more detail. A scoring step then matches the query tokens against the document tokens.

The model processes wider document context before pooling smaller chunks. This helps a chunk retain information that appears outside its own text span.

Serving capacity, latency, and modality preprocessing affect the limits of a shared API even when the model architecture supports a longer sequence.

A model or processing path may need to warm up. Compare sustained latency separately from the first request and handle timeouts in your application.

Smaller vectors reduce storage and transfer costs. For models supporting Matryoshka dimensions, evaluate how much retrieval quality is retained at your chosen size.

Limits differ by input format and hosted model. Prefer accessible URLs for large files and consult the selected model’s schema before uploading.

Some hosted models may need to load or initialize before serving a request. Latency also depends on the input size and current service load. Keep a timeout suited to the workload and inspect errors before retrying.

A key can access the services assigned to its scopes. Review the key in the dashboard before sharing it with an application.

Open the key manager to inspect recorded requests, status, latency, and available usage data for the selected key.

Sign in to manage your keys. If a key may be exposed, revoke it and update the applications that use it.

Access follows the expiry and status configured for the key. A key that is revoked or expired cannot be used for new requests.

Use only the transfer options available in your account. Do not assume that provider balances and Chatsax balances are interchangeable.

Use the key manager. Revocation stops future access, so update any dependent applications before retiring a key.

Check the platform privacy terms and the selected provider’s policy. Avoid sending material your organization has not approved for that service.

Use the status and retry information returned by the API. Add backoff and control concurrency instead of continuously retrying the same request.

The batch workflow supports submitting a job, checking its status, and downloading its output when the selected provider completes it.

Chatsax is independently operated. The current model APIs are provided by Jina, whose papers and model cards are linked for reference.

Each service has its own metering rules. Reading, search, embedding, and generation can count different units; review the active pricing before use.

Open the dashboard to receive a visitor key automatically. Sign in when you need account management or token purchases. The account screen shows your available balance and access.

The provider may have processed work before the connection ended. Review the recorded usage and billing status rather than inferring them from a closed browser tab.

The checkout screen displays payment methods available for your region, currency, and account.

Use the billing portal linked from your account, and retain the payment confirmation associated with the purchase.

No. API credit pays for hosted usage. Downloading and commercially running model weights follows the model’s separate license.

Use the contact page to describe the services, traffic, latency, and deployment requirements you need.

Check the selected account and key, the payment status, and the latest usage records. Keep the payment reference when contacting support.

Use the billing controls provided in the account or payment portal. Check the confirmation after changing an automatic payment setting.