conversationSummary()

Auto-summarizes a conversation when token count exceeds a threshold.

Signature

function conversationSummary(opts: ConversationSummaryOptions): ConversationSummaryStore

Parameters

Parameter	Type	Description
opts	ConversationSummaryOptions	Chat stream, LLM store, and token threshold.

Returns

ConversationSummaryStore — Store<string> with the rolling summary, plus destroy().

Basic Usage

import { state } from 'callbag-recharge';
import { chatStream, conversationSummary, fromLLM, ragPipeline } from 'callbag-recharge/ai';

const llm = fromLLM({ provider: 'ollama', model: 'llama4' });
const chat = chatStream(async function* (messages, signal) { ... });
const summary = conversationSummary({ chat, llm, maxTokens: 2000 });

const query = state('');
const rag = ragPipeline({ query, docSearch, llm, summary });
// summary.get() → rolling summary injected as SUMMARY section in rag.context

Options / Behavior Details

• Trigger window: Only fires after an assistant response (last message is "assistant"), when the LLM is idle (status !== "active").
• Rolling: Each summarization replaces the previous summary. Chain with systemPromptBuilder or pass as summary to ragPipeline to inject into the system prompt.
• Shared LLM: Safe to share llm with ragPipeline. The status !== "active" guard prevents summarization from interrupting an in-progress generation.