mirror of
https://gitee.com/mateos/mateclaw.git
synced 2026-09-17 04:44:40 +08:00
76 lines
3.1 KiB
Java
76 lines
3.1 KiB
Java
package vip.mate.config;
|
||
|
||
import lombok.Data;
|
||
import org.springframework.boot.context.properties.ConfigurationProperties;
|
||
|
||
/**
|
||
* 会话历史上下文窗口管理配置
|
||
*
|
||
* @author MateClaw Team
|
||
*/
|
||
@Data
|
||
@ConfigurationProperties(prefix = "mate.agent.conversation.window")
|
||
public class ConversationWindowProperties {
|
||
|
||
/** 全局默认最大输入 token（上下文窗口） */
|
||
private int defaultMaxInputTokens = 128000;
|
||
|
||
/** 历史 token 占比达此阈值触发压缩（0-1） */
|
||
private double compactTriggerRatio = 0.75;
|
||
|
||
/** 压缩后保留最近 N 轮对话（user+assistant 算一轮） */
|
||
private int preserveRecentPairs = 5;
|
||
|
||
/** 摘要自身最大 token 数（仅作 LLM maxToken 参数上限，实际预算由动态计算） */
|
||
private int summaryMaxTokens = 800;
|
||
|
||
// ==================== 动态压缩配置 ====================
|
||
|
||
/** 尾部保护的最小消息数（即使 token 预算用完也至少保留这么多） */
|
||
private int protectLastMinMessages = 10;
|
||
|
||
/** 摘要 token 预算占被压缩内容的比例 (0-1)。被压缩内容越多，摘要越长。 */
|
||
private double summaryBudgetRatio = 0.20;
|
||
|
||
/** 摘要 token 预算上限（字数，非 token） */
|
||
private int summaryBudgetCeiling = 3000;
|
||
|
||
/** 摘要 token 预算下限（字数） */
|
||
private int summaryBudgetFloor = 500;
|
||
|
||
/**
|
||
* Minimum prefix size (in messages) the pair-safe boundary must leave
|
||
* before compaction is allowed to run. After enforcing tool-call/response
|
||
* pair integrity the boundary may collapse so far forward that only a
|
||
* handful of messages remain in the prefix — at that point the
|
||
* compaction cost (a structured-summary LLM call) outweighs any token
|
||
* savings, and we may as well skip this turn.
|
||
*
|
||
* <p>Default 2 means "at least two old messages worth condensing".
|
||
* Set to 0 to always attempt compaction whenever a pair-safe cut exists.
|
||
*/
|
||
private int pairSafeMinPrefixToCompact = 2;
|
||
|
||
/**
|
||
* After compaction, re-inject the first user message of the compressed
|
||
* prefix so the original goal stays anchored in the prompt even when a
|
||
* long task has paged through dozens of turns. Without an anchor the
|
||
* structured summary alone can drift, and the model may forget what was
|
||
* being asked. Injected as a {@link org.springframework.ai.chat.messages.UserMessage}
|
||
* (never SystemMessage) so historical user input cannot be promoted to a
|
||
* system-level instruction.
|
||
*/
|
||
private boolean firstUserAnchorEnabled = true;
|
||
|
||
/**
|
||
* Maximum tokens the anchor body is allowed to consume in the prompt.
|
||
* The first user message is often short ("write me a CLI tool that…"),
|
||
* but power users sometimes paste multi-KB specs. When the body fits
|
||
* the budget it stays verbatim; when it is up to 3× over, it is
|
||
* head+tail truncated to this budget; when it is more than 3× over,
|
||
* it degrades to a 200-char pointer line so the model still knows the
|
||
* original goal existed without blowing prompt-cache or summary budget.
|
||
*/
|
||
private int firstUserAnchorMaxTokens = 400;
|
||
}
|