reworkd/AgentGPT

✨ Investigate the best similarity score threshold to remove duplicate tasks

Open

#729 建立於 2023年6月6日

在 GitHub 查看
 (0 留言) (0 反應) (0 負責人)TypeScript (34,594 star) (9,446 fork)batch import
enhancementhelp wanted

描述

When we generate tasks, we filter tasks that have a similarity score that is too close to existing tasks in the vector database

similar_tasks = memory.get_similar_tasks(
    task, score_threshold=0.95  # TODO: Once we use ReAct, revisit
)

This is done with the help of the code above. Arbitrarily, we use 0.95. Even with this, the task may not be related.

On the other hand, there may be very related / duplicated tasks that have a score that is less than this.

This ticket is tasked with investigating what the best value for this threshold is, or to use some other means of calculating similarity for this given case.

貢獻者指南

技術棧
typescript
領域
aidatabackend
議題類型
research
難度面向新貢獻者的預計實作難度，1 表示很小改動，5 表示專家級工作。
3
預計時間有經驗貢獻者完成調查、實作、測試並準備 pull request 的粗略時間範圍。
half day
活動狀態議題目前的可參與程度：新鮮、活躍、陳舊、阻塞或等待維護者輸入。
stale
清晰度議題是否清楚說明預期改動、驗收標準和下一步。
clear
前置要求
similarity metricsvector databasesAgentGPT codebase
新手友善度1-100 的估計分數，表示該議題對首次貢獻者的友善程度。
60
研究方向
Investigate the function memory.get similar tasks in the codebase to understand current implementation. Experiment with different score threshold values (e.g., 0.9, 0.8, 0.7) and evaluate the impact on task deduplication using a test dataset. Analyze the trade offs between false positives and false negatives and propose a recommended threshold or alternative similarity approach.