Inference
Chat, Messages, Responses, completions, embeddings, moderation and rerank calls. Each call goes through the request loop of proxium: authentication, budget, routing, failover and cost. The response cache applies to chat and Messages calls only.
List batches
Gives the batches of the project, newest first, with no results. `has_more` is true when another page exists.
Create a batch
Sends a batch of chat completion requests to the batch API of the vendor that `model` names: OpenRouter, OpenAI or Anthropic. The vendor answers within 24 hours at its batch price. The batch uses the project's own key for the vendor when the project has one, else the platform key. The caps of the project count the batch at its estimate before proxium sends it.
Retrieve a batch
Gives a batch of the project. proxium asks the vendor for the state of an open batch first. When the vendor ended the batch, `results` holds one line for each request in the shape of an OpenAI batch output line: `{custom_id, response: {status_code, body}, error}`.
Cancel a batch
Asks the vendor to stop a batch. The status changes to `cancelling`, and to `cancelled` when the vendor stops. The lines that finished before the stop have results and are billed. OpenRouter has no cancel for a batch.
Create a chat completion
Runs an OpenAI chat completion request through the full proxium loop: authentication, memory, the response cache, the budget, routing, failover and cost. A cache hit costs no call slot. The answer is the provider's body without change. With `stream: true`, the answer is server-sent events. A cache hit is sent as events too.
Classify a text with caller-defined checks
Answers yes/no checks of one text with Jev, through the proxium loop: authentication, the budget, routing and cost. Each check gets a probability. The answer names the flagged check with the highest probability as `classification`, and lists every flagged check in `flags`. The checks are the presets of proxium (`checks`) and the caller's own questions (`questions`). The feature is granted by plan.
Create a completion
Sends a legacy OpenAI completions request through the proxium loop: authentication, the budget, routing, failover and cost. With `stream: true`, the answer is server-sent events. The response cache does not apply to this route.
Run a saved decision
Runs a saved decision of the source project: the input, the memory and the constraints of the project build the state of the document, Jev answers the questions of the document, and the rules turn the answers into the actions of the decision. The call is a System One call under a document, so the platform Jev of the call needs the classify grant of the project.
Report what a decision call led to
Records the outcome of one call of a saved decision: the right answer for each question that the call asked, or the action taken and its result. The outcome is stored once: a second report for the same `decision_id` is refused. No model runs on this route, so the route needs no Jev grant.
Create embeddings
Sends an OpenAI embeddings request through the proxium loop: authentication, the budget, routing, failover and cost.
Create a message (Anthropic)
Accepts an Anthropic Messages request and runs it through the same proxium loop as `/v1/chat/completions`: proxium translates the request to the OpenAI chat shape, routes it to any provider, and translates the answer back to the Anthropic shape. With `stream: true`, the answer is Anthropic server-sent events. Each error body is in the Anthropic shape. The virtual key goes in `Authorization: Bearer`, or in `x-api-key` as the Anthropic SDKs send it.
Create a moderation
Checks text against the moderation categories of OpenAI, through the proxium loop: authentication, the budget, routing, failover and cost. A route that starts with a Jev model is answered by Jev, with one score for each category, in the OpenAI moderation shape. Another model must be a moderation model of a vendor that serves `/moderations`; proxium sends it the body without change.
Rerank documents
Orders documents by how well they answer a query, through the proxium loop: authentication, the budget, routing, failover and cost. A route that starts with a Jev model is answered by Jev, with one relevance score for each document, in the shape that Cohere and Jina return. Another model must be a rerank model of a vendor that serves `/rerank`; proxium sends it the body without change.
Create a model response
Sends an OpenAI Responses request through the proxium loop: authentication, memory, the budget, routing, failover and cost. With `stream: true`, the answer is server-sent events. The response cache does not apply to this route.
Send a call to Jev as it is
Forwards a System One call to Jev through the proxium loop: authentication, the guardrails, the budget, routing and cost. The body and the answer are Jev's own, so a System One client changes only its base URL and its key. Proxium runs no action after the answer. A repeated call is answered from the project's exact cache. The request and the answer of each call are kept by default; a project turns that off in its settings. The feature is granted by plan, as `/v1/classify` is.