Summarize messages

summarizeMessages() is a pure function: pass in messages, get back a summary string and stats. No store, no eviction, no policy — useful when you have a bounded batch of messages and want one summary right now.

What problem does this solve?

Sometimes you don't want a stateful, automatic compaction pipeline. You have a finished conversation, a closed thread, or a one-off batch — and you need a summary you can store, log, or feed to another prompt. summarizeMessages() is the smallest unit that does that.

When should I use it?

• You have a bounded batch of messages and want one summary
• You're processing a finished conversation (e.g. for archival or a digest email)
• You want to manually compact at a moment of your choosing — no auto-eviction

When should I NOT use it?

• You're building a chat with rolling history — that's sliding window, which calls this internally
• You need structured output — use extract key facts
• You want per-turn pressure tracking — use budget manager

Quick start

import { summarizeMessages } from '@crux/core/compaction'
import { generateTextFn } from '@crux/ai'

const result = await summarizeMessages({
  messages: conversationHistory,
  generate: generateTextFn,
  model: cheapModel,
  maxTokens: 500,
  focus: ['decisions', 'action-items'],
})

result.summary       // string
result.tokensBefore  // original token count
result.tokensAfter   // summary token count
result.ratio         // compression ratio (e.g. 0.15 = 85% reduction)

Steering with focus

The focus parameter is an array of topic keywords. The summarizer prioritizes information related to those topics:

focus: ['decisions', 'action-items']     // project conversations
focus: ['user-preferences', 'questions'] // support conversations
focus: ['blockers', 'risks']             // status updates

Common values: decisions, action-items, questions, agreements, blockers, risks, user-preferences, open-questions.

You can also pass a system override to the summarizer for full control:

await summarizeMessages({
  messages,
  generate: generateTextFn,
  model: cheapModel,
  system: 'You are summarizing a customer support conversation. Preserve customer concerns verbatim.',
})

Where to next