Struct LlmSummarizer 

pub struct LlmSummarizer { /* private fields */ }

Summarizes older conversation turns by asking DeepSeek to write a concise prose summary, then replaces the compressed turns with a single Role::System message containing that summary.

Trigger

Fires when the estimated token count of the compressible portion of the history (everything except permanent system prompts) exceeds token_threshold.

Behavior

1. Permanent Role::System messages (user-provided via with_system_prompt) are extracted and re-prepended after summarization — they are never lost.
2. Any previous [auto-summary] system message is included in the text sent to the model so the new summary is cumulative.
3. The retain_last most recent non-system turns are kept verbatim; everything older is replaced by the LLM-generated summary.
4. If the API call fails the history is left unchanged and the error is returned so the caller can decide whether to abort or continue.

Example

use ds_api::{DeepseekAgent, ApiClient};
use ds_api::conversation::LlmSummarizer;

let summarizer = LlmSummarizer::new(ApiClient::new("sk-..."));
let agent = DeepseekAgent::new("sk-...")
    .with_summarizer(summarizer);

Implementations

impl LlmSummarizer

pub fn new(client: ApiClient) -> Self

Create with default thresholds: trigger at ~60 000 tokens, retain last 10 turns.

pub fn token_threshold(self, n: usize) -> Self

Builder: set a custom token threshold.

pub fn retain_last(self, n: usize) -> Self

Builder: set how many recent messages to keep verbatim.

Trait Implementations

impl Clone for LlmSummarizer

fn clone(&self) -> LlmSummarizer

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Summarizer for LlmSummarizer

fn should_summarize(&self, history: &[Message]) -> bool

Return true if the history should be summarized before the next API turn.

fn summarize<'a>( &'a self, history: &'a mut Vec<Message>, ) -> Pin<Box<dyn Future<Output = Result<(), ApiError>> + Send + 'a>>

Compress history in-place, returning an error only for unrecoverable failures.