Skip to main content

Prompt Pipeline

The prompt pipeline is the fixed, ordered set of stages that turns a chat into a provider request, from user input to the saved message.

What the Pipeline Does​

Every generation — a new message, a swipe, a regeneration, or an impersonation — runs through the same stages in the same order. The pipeline assembles the context from the character, persona, lorebook, and memory, counts tokens, fits the context into the model's budget, lets plugins intercept, renders the request in the selected instruct format, and finally streams and saves the response.

Pages in This Section​

  • Pipeline Stages — the 14 stages in order and the rules every plugin hook must follow.
  • Instruct Formats — how the clean message array is rendered with sandboxed Handlebars templates.
  • Tokenization — the local tokenizer registry and its approximate fallback.
  • Context Shifting — how the pipeline fits the context into the token budget and which strategies exist.

Implementation​

The pipeline lives in apps/server/src/pipeline/. It runs entirely on the server, before any network call, so the request that reaches a provider is always the result of the same deterministic stages.

  • Plugin interceptors and their registration APIs are documented in the Plugin SDK.
  • The generation endpoint and the context audit are part of the API Reference.
  • Provider adapters that consume the serialized request are documented under Providers.