Repository navigation
增加可选时间注释,保留原文检索证据并支持离线加工对照 - #25
Merged
Merged
Conversation
Contributor
Author
|
QA 部署进展:按用户针对本次的明确授权,仅通过 kubectl 替换 mb adapter 镜像,其他 Pod 配置及数据盘保持不变。源码提交 公开 LoCoMo 历史加工及对照任务已启动:10 份历史、861 道文本题。固定原文检索上下文及旧回答,派生候选加入后重新回答,两组均使用完整 Refined Judge 提示词评分;所有失败留档后继续。结果待补充。 PR 保持未合并;CI 预算问题未改动。 |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
本 PR 提供记忆加工实验入口;最新迭代收窄为默认关闭的时间注释,避免宽泛事实提取和固定派生配额干扰原文。
当前 Search 路径通过 MEMORY_TIME_NORMALIZATION_ENABLED=true 启用:仅对有明确消息时间锚点的相对时间附加确定性的日历解释,不调用加工 LLM,不新增数据库写入或检索。原文、记录 ID、条数和排序保留;缺少锚点、日期冲突、引文、续接分片及未支持表达直接跳过。每条和全响应均限制注释大小,预算耗尽不会删除原文。Add 和 journal 协议不变,关闭开关即可回退。
英文 last Friday、中文上周五及周范围的解释约定在注释和文档中明确标记,不视为推断的新事件事实。旧离线 LLM CLI 仅保留用于复现先前实验,与当前时间开关独立。
验证:621 项测试通过,Ruff 通过;覆盖时间边界、引文、缺失/冲突锚点、Add 幂等性,以及 100 条 Search 结果在全局注释预算耗尽后仍全部保留。新增镜像已构建;本地冻结检索候选的完整配对对照已完成。仅使用公开/合成材料,不包含官方测试内容或环境凭据。
本地冻结检索的完整同轮配对对照:原文 653/861(75.84%),时间注释 680/861(78.98%),改善 52 题、退化 25 题,技术失败/未评分 0 题。两组均重新回答并采用完整官方 Refined Judge prompt;模型配置保持一致,但不等同官方成绩。每题原有 100 条证据均保留。该结果是本地新镜像重放,不是新一轮 QA 端到端压测。
新增镜像内 HTTP 冒烟通过:健康检查、认证、Add/Search、开关切换后的幂等性及无时间内容保持原样。当前 PR CI 已通过,PR 保持未合并。
本轮限制:仍有 25 题退化;回答模型平均输入 token 增加约 12.3%。结果只支持本轮冻结检索的 LoCoMo 对照,不证明其他 benchmark 或真实并发链路同样受益。