Struct ContextWindow 

pub struct ContextWindow<M: BaseChatModel> { /* private fields */ }

Context window manager for fitting messages within a token budget.

Counts tokens using a TokenCounter (defaults to TiktokenCounter), and applies a Strategy when messages exceed max_tokens.

Implementations

impl<M: BaseChatModel> ContextWindow<M>

pub fn new(max_tokens: usize) -> Result<Self, MemoryError>

Creates a new ContextWindow with the Truncate strategy and default TiktokenCounter.

P1-4: returns Result instead of panicking — a tiktoken model download/load failure (offline / missing model) returns MemoryError, so the library constructor no longer crashes on a bad local environment.

pub fn with_strategy( max_tokens: usize, strategy: Strategy<M>, ) -> Result<Self, MemoryError>

Creates a new ContextWindow with a specific strategy and default TiktokenCounter.

pub fn with_max_tokens(max_tokens: usize) -> Result<Self, MemoryError>

Creates a new ContextWindow with a custom max token limit and default counter/strategy.

pub fn with_counter(self, counter: Arc<dyn TokenCounter>) -> Self

Sets a custom token counter.

pub fn max_tokens(&self) -> usize

Returns the max token limit.

pub async fn fit( &self, messages: Vec<Message>, ) -> Result<Vec<Message>, MemoryError>

Fits messages within the token limit by applying the configured strategy.

• If total tokens are within max_tokens, returns messages as-is.
• If over, applies the Strategy (truncate or summarize).

Budget semantics (P1-4 contract)

Strategy::Truncate: System messages are always kept and do not consume budget (M7); if the System messages alone exceed max_tokens, they are returned as-is (the result may exceed the budget). Even when the budget is too small for a single conversation message, at least the newest non-System message is kept — history is never silently emptied (H7). Callers must not assume fit’s result is always within budget.

Strategy::Summarize: the summary placeholder counts toward the budget, but the LLM’s actual summary token count is unknown — the budget semantics differ from Truncate, and the intentionally inconsistent handling of System messages between the two is deliberate; each strategy defines its own.

Arguments

• messages - The conversation messages to fit.

Returns

A vector of messages that fits within the token budget.