What proxium does not do
This page lists what proxium.tech does not do. For each gap, it says what to do instead, if anything.
No speech to text
proxium has no route for transcription or translation of audio. /v1/audio/speech is the other direction: it turns text into speech. /v1/audio/transcriptions and /v1/audio/translations do not exist.
Instead: send speech-to-text calls to the vendor directly. proxium does not record their cost.
No content screening on proxium.tech
proxium.tech does not redact or block personal data or secrets in your requests. It sends the body to the provider as you send it.
Instead: remove sensitive data in your application before you call proxium.
Data stays until you delete it
proxium keeps the attempt records, the usage, the stored prompts and answers, and the memories until a person deletes them. Two things expire on their own:
- Memory conversations and closed memories, when the project sets Keep conversations for.
- Cached answers, after 3600 seconds.
Instead: in Settings, set Stored prompts and answers to the narrowest value you need. Keep conversations for is in the Memory section of Settings. To delete data, erase one end user or erase the project. Data handling explains both.
/v1/models lists the models of the project, not of the key
/v1/models returns every provider/model id that the project can route to. A virtual key can have its own list of allowed models. /v1/models does not apply that list.
Instead: keep the list of allowed models of each key in your application. A call with a model that the key does not allow gets 400.
No response cache on most routes
The response cache serves /v1/chat/completions and /v1/messages only. These routes send every request to a provider:
| Route | Response cache |
|---|---|
/v1/responses | No |
/v1/embeddings, /v1/moderations, /v1/rerank | No |
/v1/images/generations, /v1/audio/speech | No |
/video/generations, /video/operations, /video/download | No |
Instead: if repeated requests cost you money, send them as chat or Messages calls. The response cache explains how it works.
No way to skip the cache for one call
No request header makes proxium skip the cache. A project cannot turn the cache off. The answer does not say that it came from the cache.
Instead: change the body. The cache key holds every field of the body except stream, stream_options and user. A request that differs in another field misses the exact layer. It can still match a near request in the semantic layer. The Cache panel on Requests counts the hits.
Some routes have no /v1 form
These routes are at the origin, https://proxium.tech, not under https://proxium.tech/v1:
| Route | Use |
|---|---|
/video/generations | Start a video |
/video/operations | Poll a video |
/video/download | Download a finished video |
/status/peek | Read the budget state before a call |
/usage/me | Read the spend of the project this month |
Instead: call them with the full origin URL, for example https://proxium.tech/usage/me.
No failover and no timeout header for images and speech
/v1/images/generations and /v1/audio/speech call one provider. They do not fail over to another provider, and they do not read x-proxium-timeout-ms.
Instead: if an image or speech call fails, retry it in your application.
Messages requests lose some fields
proxium translates an Anthropic Messages request to the chat format, so it can route it to any provider. It translates these fields: system, messages, tools, tool_choice, temperature, top_p, top_k, stop_sequences, stream and metadata.user_id. It drops all other fields. It also drops thinking, redacted_thinking and document blocks.
Instead: if you need extended thinking or document blocks, send the call to the vendor directly.
Memory keeps conversations only
Memory keeps only chat, Messages and Responses calls. On /v1/embeddings, /v1/moderations and /v1/rerank, a valid x-proxium-memory value has no effect. recall puts memories into chat and Messages calls only. On /v1/responses, recall acts as write.
Instead: use /v1/chat/completions or /v1/messages when a call needs the memories of the project. Use project memory explains the modes.