検索設計から 推論設計への重心移動と Recall-First Retrieval

by po3rin

Embed

Start on current slide

Slide 1

Slide 1 text

ݕࡧઃܭ͔Β ਪ࿦ઃܭ΁ͷॏ৺Ҡಈͱ 3FDBMM'JSTU3FUSJFWBM

Slide 2

Slide 2 text

)JSPNV/BLBNVSB )JSPNV /BLBNVSB QPSJO "*.-0QT&OHJOFFS *OUFSFTUJOHJO *OGPSNBUJPO3FUSJFWBM /BHPZB6OJW . *OD -BZFS9*OD (SBEVBUF4DIPPMPG4DJFODF "*ɾ.-5FBN.-0QT%BUB&OHJOFFSJOH "*69.-0QT5FDI-FFE "*69.-0QT5FDI-FFE d

Slide 3

Slide 3 text

1 ݕࡧઃܭ͔Βਪ࿦ઃܭ΁ͷॏ৺Ҡಈͱ3FDBMM'JSTU3FUSJFWBM TODAY'S THESIS RAG ͷ඼࣭վળͷॏ৺͸ɺ ʮݕࡧઃܭʯ͔Β ʮਪ࿦ઃܭʯ΁ҠΓͭͭ͋Δɻ ͦͷస׵ͷഎܠɾख๏ɾͦͯ͠ Recall-First Retrieval ͱ͍͏ઃܭࢥ૝Λ੔ཧ͠·͢ɻ

Slide 4

Slide 4 text

2 ݕࡧઃܭ͔Βਪ࿦ઃܭ΁ͷॏ৺Ҡಈͱ3FDBMM'JSTU3FUSJFWBM

Slide 5

Slide 5 text

ݕࡧઃܭͷຽओԽͱɺͦͷนɻ ݕࡧઃܭ͔Βਪ࿦ઃܭ΁ͷॏ৺Ҡಈͱ3FDBMM'JSTU3FUSJFWBM

Slide 6

Slide 6 text

3 ݕࡧઃܭ͔Βਪ࿦ઃܭ΁ͷॏ৺Ҡಈͱ3FDBMM'JSTU3FUSJFWBM RAGͷ  ࣮૷͸ඪ४Խ͞Εͨɻ CHAPTER 1 · DEMOCRATIZATION

Slide 7

Slide 7 text

4 ݕࡧઃܭ͔Βਪ࿦ઃܭ΁ͷॏ৺Ҡಈͱ3FDBMM'JSTU3FUSJFWBM ͔͠͠ݕࡧͷνϡʔχϯά͸ऴΘΒͳ͍ɻ CHAPTER 1 · THE WALLS

Slide 8

Slide 8 text

ݕࡧઃܭ͔Βɺਪ࿦ઃܭ΁ ݕࡧઃܭ͔Βਪ࿦ઃܭ΁ͷॏ৺Ҡಈͱ3FDBMM'JSTU3FUSJFWBM

Slide 9

Slide 9 text

6 ݕࡧઃܭ͔Βਪ࿦ઃܭ΁ͷॏ৺Ҡಈͱ3FDBMM'JSTU3FUSJFWBM • LLM ͕े෼ڧ͚Ε͹ɺݕࡧଆ͕׬ᘳͰͳͯ͘΋࠷ऴճ ౴ͷ࣭͸อͨΕΔɻ • BM25 ͷΑ͏ͳݹయతΩʔϫʔυݕࡧͰ΋ɺจॻηο τ͕ἧ͑͹ LLM ͕ิ׬͢Δɻ • hybrid + rerank ͷΑ͏ʹݕࡧଆΛް͘͢Δඅ༻ରޮՌ ͸LLM ͕ڧ͘ͳΔ΄ͲബΕΔɻ Keyword Search  Is All You Need CHAPTER 2 · EVIDENCE ڧྗͳ LLM Λલఏʹ͢Ε͹ɺBM25 ʴ LLM ͚ͩͰ hybrid + rerank ʹඖఢ͢Δ඼࣭͕ग़Δ͜ͱΛ࣮ূ [1]ɻ

Slide 10

Slide 10 text

7 ݕࡧઃܭ͔Βਪ࿦ઃܭ΁ͷॏ৺Ҡಈͱ3FDBMM'JSTU3FUSJFWBM CHAPTER 3 · CONDITIONS ॏ৺͕ಈ͘3ͭͷ৚݅ɻ

Slide 11

Slide 11 text

8 ݕࡧઃܭ͔Βਪ࿦ઃܭ΁ͷॏ৺Ҡಈͱ3FDBMM'JSTU3FUSJFWBM CHAPTER 3 · The Bitter Lesson RICH SUTTON · 2019 ʮ௕ظతʹ͸ɺ൚༻తͳܭࢉΛ εέʔϧͤ͞Δํ๏͕ɺ ਓ͕ؒखͰ૊ΈࠐΜͩ஌ࣝΛӽ͑Δɻʯ The Bitter Lesson [2]

Slide 12

Slide 12 text

ਪ࿦ઃܭͷͭͷख๏ɻ ݕࡧઃܭ͔Βਪ࿦ઃܭ΁ͷॏ৺Ҡಈͱ3FDBMM'JSTU3FUSJFWBM

Slide 13

Slide 13 text

9 ݕࡧઃܭ͔Βਪ࿦ઃܭ΁ͷॏ৺Ҡಈͱ3FDBMM'JSTU3FUSJFWBM

Slide 14

Slide 14 text

10 ݕࡧઃܭ͔Βਪ࿦ઃܭ΁ͷॏ৺Ҡಈͱ3FDBMM'JSTU3FUSJFWBM

Slide 15

Slide 15 text

11 ݕࡧઃܭ͔Βਪ࿦ઃܭ΁ͷॏ৺Ҡಈͱ3FDBMM'JSTU3FUSJFWBM

Slide 16

Slide 16 text

12 ݕࡧઃܭ͔Βਪ࿦ઃܭ΁ͷॏ৺Ҡಈͱ3FDBMM'JSTU3FUSJFWBM • Recall-FirstͰͱ͖ͬͯͨυΩϡϝϯτΛਫ਼ಡ͢ΔͨΊͷίϯϙʔωϯτͱͯ͠ར༻͢Δ ࠷ۙ͸LLM WikiͳͲNavigateϕʔεͷख๏ͷఏҊ͕૿͖͑ͯͨ

Slide 17

Slide 17 text

3FDBMM'JSTU3FUSJFWBM ݕࡧઃܭ͔Βਪ࿦ઃܭ΁ͷॏ৺Ҡಈͱ3FDBMM'JSTU3FUSJFWBM

Slide 18

Slide 18 text

13 ݕࡧઃܭ͔Βਪ࿦ઃܭ΁ͷॏ৺Ҡಈͱ3FDBMM'JSTU3FUSJFWBM

Slide 19

Slide 19 text

14 ݕࡧઃܭ͔Βਪ࿦ઃܭ΁ͷॏ৺Ҡಈͱ3FDBMM'JSTU3FUSJFWBM TRADE-OFFS ͨͩ͠ɺ ۜͷ஄ؙͰ͸ͳ͍ɻ

Slide 20

Slide 20 text

1 ݕࡧઃܭ͔Βਪ࿦ઃܭ΁ͷॏ৺Ҡಈͱ3FDBMM'JSTU3FUSJFWBM ݕࡧઃܭ͔Βਪ࿦ઃܭ΁ͷॏ৺Ҡಈͱ3FDBMM'JSTU3FUSJFWBM

Slide 21

Slide 21 text

16 ݕࡧઃܭ͔Βਪ࿦ઃܭ΁ͷॏ৺Ҡಈͱ3FDBMM'JSTU3FUSJFWBM ͱ͍͏࿩Λٕज़ॻయͰॻ͖·ͨ͠ https://techbookfest.org/product/jbe1zdnjqbb6cHwebRhapn?productVariantID=3VefBkLxtiAHJr1vAHjJMY

Slide 22

Slide 22 text

17 ݕࡧઃܭ͔Βਪ࿦ઃܭ΁ͷॏ৺Ҡಈͱ3FDBMM'JSTU3FUSJFWBM References 1. Keyword Search Is All You Need (AmazonͷݚڀνʔϜʹΑΔൃද) 1. URL: https://www.amazon.science/publications/keyword-search-is-all-you-need-achieving-rag-level-performance-without-vector-databases- using-agentic-tool-use 2. Bitter Lesson 1. http://incompleteideas.net/IncIdeas/BitterLesson.html 3. Search-o1 (Agentic Searchͷ୅දతͳ࣮૷) 1. arXiv: https://arxiv.org/abs/2501.05366 4. SPD-RAG (Sub-agent Per Document RAG) 1. arXiv: https://arxiv.org/abs/2603.08329 5. PageIndex (Vectorless RAGͷ۩ମతͳ࣮૷) 1. URL: https://pageindex.ai/ 6. SRLM (Self-Re fl ective Program Search) 1. arXiv: https://arxiv.org/abs/2603.15653

Slide 23

Slide 23 text

ݕࡧઃܭ͔Β ਪ࿦ઃܭ΁ͷॏ৺Ҡಈͱ 3FDBMM'JSTU3FUSJFWBM