Planck.AI.Context (Planck.AI v0.2.2)

Everything sent to the LLM in a single request: system prompt, conversation history, and available tools.

Inference parameters (temperature, max_tokens, etc.) are NOT stored here — they are passed as keyword options at the Planck.AI.stream/3 or Planck.AI.complete/3 call site and forwarded directly to req_llm.

Examples

iex> %Planck.AI.Context{
...>   system: "You are a helpful coding assistant.",
...>   messages: [
...>     %Planck.AI.Message{role: :user, content: [{:text, "Hello"}]}
...>   ],
...>   tools: []
...> }

Summary

Types

t()

Functions

estimate_tokens(context)

Rough token estimate for the entire request this context represents — system prompt, conversation, and tool schemas.

Types

t()

@type t() :: %Planck.AI.Context{
  messages: [Planck.AI.Message.t()],
  system: String.t() | nil,
  tools: [Planck.AI.Tool.t()]
}

Functions

estimate_tokens(context)

@spec estimate_tokens(t()) :: non_neg_integer()

Rough token estimate for the entire request this context represents — system prompt, conversation, and tool schemas.

A live "how much context is used" figure needs all three: the system prompt alone is routinely the largest single piece (tool guidance, skills, AGENTS.md), and tool schemas sent on every request can be substantial too once several tools are registered — estimating only the conversation messages undercounts by everything actually sent alongside them. This is the actual t() about to be (or just was) sent, not a reconstruction of it from separate pieces that can drift out of sync with each other. This is also the one place per-content-part token counting is written — a caller holding Planck.Agent.Message.t() structs converts via Planck.Agent.Message.to_ai_messages/1 and wraps the result in a bare %__MODULE__{} (system/tools left at their defaults for a messages-only estimate) rather than duplicating this logic for its own message type.