Context Retrieval
Retrieval Evaluation
Retrieval Dataset v1.0.0 · Agent Dataset v1.1.0。在相同 Ground Truth 與相同 K 下比較 Keyword 與 Hybrid Semantic Retrieval。
目前結論
待審核品質持平,Hybrid 增加延遲
目前小型 fixture 中 Keyword 已達到 100% Recall,沒有證據支持全面改用 embedding。
Keyword
Recall@K100%
Precision@K40%
MRR0.90
平均延遲<1 ms
Context333 chars
vs
Hybrid Semantic
Recall@K100%
Precision@K40%
MRR0.90
平均延遲1358 ms
Context328 chars
逐案例檢索結果
Ground Truth 是每題解題或判斷風險所需的檔案;Patch 成功率不混入檢索分數。
| 案例 | Ground Truth | 文件數 | Keyword 排名 | Hybrid 排名 | Recall |
|---|---|---|---|---|---|
| 修正 cancelled 狀態顏色patch-cancelled-badge | status-badge.js | 20 |
|
| 100% / 100% |
| 修正分頁少一筆patch-pagination-boundary | pagination.js | 20 |
|
| 100% / 100% |
| 修正 Email 正規化patch-normalize-email | normalize-email.js | 20 |
|
| 100% / 100% |
| 跨檔案傳遞發票幣別patch-invoice-currency | currency.jsinvoice.js | 21 |
|
| 100% / 100% |
| 庫存不可降為負數patch-inventory-floor | inventory.js | 20 |
|
| 100% / 100% |
限制與正確解讀
- Each measured case ranks 19 to 21 files, including a shared corpus with overlapping names, keywords, and responsibilities.
- Patch success is reported separately because model behavior is a confounding variable.
- 政策拒絕與資訊不足案例沒有可由 Query 推導的目標檔案,因此保留在 Agent Benchmark,但不製造假的 Retrieval Ground Truth。