Message Trimmer Demo

Demonstrating trim functions on real WildChat-1M conversations

About This Demo

This demo shows how the trimMiddle() function works in the LLM router architecture. The function trims messages by keeping the start (60%) and end (40%) with an ellipsis in between, reducing token count while preserving context for intent classification. Adjust the max length sliders below to see how different messages are trimmed.

Max Assistant Message Length: 500

Max Previous User Message Length: 400

Select Conversation:

Implementation (from src/lib/server/router/arch.ts)// Trim assistant messages to reduce routing prompt size
if (m.role === "assistant") {
  return { ...m, content: trimMiddle(m.content, maxAssistantLength) };
}
// Trim previous user messages (latest preserved)
if (m.role === "user" && idx !== lastUserIndex) {
  return { ...m, content: trimMiddle(m.content, maxPrevUserLength) };
}

Loading conversations...