planner work process

2025-10-07 · 2 min read · #

planner

索引产物

  • dense_coarse.jsonl / coarse.faiss:section/image/table 的语义摘要(MMR)
  • sparse_coarse.jsonl / bm25_coarse/:多域(title/caption/table_schema/labels/aliases/body)
  • dense_leaf.jsonl(仅 chunk) / leaf.faiss / bm25_leaf/:叶子级检索(Observer 用)
  • graph_edges.jsonl:child / parent / sibling / same_page / ref / has_col
  • id_maps.json:label2id、figure、table 角色:section、image(figure)、table Planner 输出计划,不是执行结果;执行交给 Observer,判定与补证由 Reasoner 完成。 LLM-Only:无 LLM 直接报错,不走规则回退。

让agents自己决定使用哪种检索方法

1 特定任务 医疗场景

2 小模型碰瓷大模型

3 可解释性,怎么证明你的可解释性比CoT优越呢

4 长文档理解,但是做攻击或隐私保护

Type        SamplesCorrectAccuracy
Single-page     370      89  24.05%
Cross-page      465      32    6.88%
Unknown         219    146  66.67%
Overall   1,054    267  25.33%
Evidence Source SamplesCorrectAccuracy
Chart               174      36  20.69%
Table               210      30  14.29%
Pure-text           300      51  17.00%
Generalized-text    115      11    9.57%
Figure              296      26    8.78%
Overall       1,054    267  25.33%

Router Hint Accuracy (Hit-on-Any Matching Evidence Source)

Router HintEvidence TargetsCountCorrectAccuracy
textPure-text (Plain-text)48417435.95 %
tableTable25310842.69 %
graphicsChart, Figure22617878.76 %
layoutGeneralized-text (Layout)23834.78 %
totalAll above98646847.46 %

Support Nodes‘ Page Accuracy

MetricValue
Hit@Any35.45 % (296 / 835)
precision0.2341
recall0.2400
F10.2223

Router Hint Accuracy (Doc)

Router HintEvidence TargetsCountCorrectAccuracy
textPure-text (Plain-text)36714740.05 %
tableTable1909550.00 %
graphicsChart, Figure15312279.74 %
layoutGeneralized-text (Layout)18633.33 %
Overall     All above77837047.58 %
Router HintEvidence TargetsBaseline CountBaseline CorrectBaseline AccuracyDocTree CountDocTree CorrectDocTree AccuracyΔ Accuracy
textPure-text (Plain-text)41316239.23 %30013946.33 %+7.11 pts
tableTable1999648.24 %29613846.62 %−1.62 pts
graphicsChart, Figure18114680.66 %23418679.49 %−1.18 pts
layoutGeneralized-text (Layout)19631.58 %9222.22 %−9.36 pts
unknown5100.00 %3200.00 %0.00 pts
none800.00 %00
TotalAll above87141047.07 %87146553.39 %+6.31 pts