Reference
Glossary
- Adapter
- The plug-in that translates between the OpenAI wire format and one provider’s protocol. Janus ships adapters for OpenAI-compatible endpoints, Google Gemini, Anthropic, AWS Bedrock, Google Vertex, Ollama, vLLM, llama.cpp, and Hugging Face TEI.
- Upstream
- A configured provider endpoint: a base URL, an adapter type, and an encrypted credential.
- Model
- One model discovered on an upstream. Discovered models stay disabled until an administrator approves and prices them.
- Grant
- Permission for a user, a group, all users, a team, all teams, or a service token to call a model. Personal access combines direct, group and all-users grants — the most permissive match wins. Working under a team uses only that team’s grants and all-teams grants.
- Rate card
- The per-million-token prices used to attribute spend. Rate cards are versioned with an effective-from date so historical usage keeps its original price.
- Modality
- What kind of work a request performs: chat, embedding, image, audio, tts, stt, video, function, response, assistant, thread, file, moderation, or fine_tune.
- Cached tokens
- Input tokens the provider served from its own prompt cache, billed at the cached rate. OpenAI-style providers count them inside tokens_in; Anthropic and Bedrock report them separately from tokens_in.
- Cache-write tokens
- Tokens written to the provider’s prompt cache (5-minute or 1-hour). They are billed at the model’s cache-write rates; providers that do not report them meter zero.
- Context window
- The maximum number of tokens a model accepts as input (context_window). 0 is shown as a dash, meaning unknown.
- Quota window
- The period a limit is measured over. Calendar windows (day, week starting Monday, month) snap to UTC boundaries; rolling windows (24 hours, 7 days, 30 days) slide continuously.
- Breach behavior
- What happens when a limit is crossed mid-request. The default lets the in-flight request finish and refuses the next one; hard-kill cuts a stream within a fraction of a second of the limit being crossed and ends it with a policy.quota_exceeded error frame. Non-streaming requests always finish.
- TTFB
- Time to first byte — how long the provider took to start responding, separate from total latency.
- Policy rule
- A composable block rule over client-supplied signals. A policy control, not a security control: every signal but the token can be spoofed.