ContextPipeline

The run-time context builder.

ContextPipeline constructs a ChatRequest from the user input and the session state. It runs a sequence of ContextAdapters, applies a compaction filter (so already-summarised messages are not re-sent), and trims the result to fit the token budget.

The full file is src/runtime/context.rs.

Stages

1. Adapters — call each ContextAdapter registered in Extensions::context_adapters to inject extra context (system prompts, RAG snippets, function outputs).
2. Compaction filter — drop messages marked is_compaction: true whose CompactionMeta::summary_text is in scope; the model receives the summary instead of the raw messages.
3. Token trim — fit the message list into the model's context window by trimming the oldest non-system messages.

API

pub struct ContextPipeline {
    adapters: Vec<Arc<dyn ContextAdapter>>,
    max_history_messages: usize,
    max_history_tokens: usize,
    enable_compaction_filter: bool,
}

impl ContextPipeline {
    pub fn new() -> Self;
    pub fn with_adapter(self, a: Arc<dyn ContextAdapter>) -> Self;
    pub fn with_max_history(self, n: usize) -> Self;
    pub fn with_max_history_tokens(self, n: usize) -> Self;
    pub fn with_compaction_filter(self, on: bool) -> Self;

    pub async fn build(
        &self,
        session: &Session,
        input: &str,
        store: &dyn SessionStore,
    ) -> Result<ChatRequest, ContextError>;
}

Worked example

let pipeline = ContextPipeline::new()
    .with_adapter(Arc::new(StaticAdapter::new("You are concise.")))
    .with_adapter(Arc::new(RagContextAdapter::new(qdrant)))
    .with_max_history(50)
    .with_max_history_tokens(64_000)
    .with_compaction_filter(true);

let req = pipeline.build(&session, "Hello", &*store).await?;

See also

• CompactionService — produces the compaction messages.
• ContextAdapter — the adapter trait.
• RAG Context Adapter — the RAG implementation.