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

Transcript

1.  ݕࡧઃܭ͔Β
   ਪ࿦ઃܭ΁ͷॏ৺Ҡಈͱ
   3FDBMM'JSTU
   3FUSJFWBM
   
2. )JSPNV
   /BLBNVSB )JSPNV
   /BLBNVSB QPSJO
   
   "*.-0QT
   &OHJOFFS
   *OUFSFTUJOH
   JO
   
   *OGPSNBUJPO
   3FUSJFWBM /BHPZB
   6OJW .
   *OD -BZFS9
   *OD (SBEVBUF
   4DIPPM
   PG
   4DJFODF "*ɾ.-
   5FBN
   .-0QT
   
   %BUB
   &OHJOFFSJOH "*69.-0QT
   5FDI-FFE "*69.-0QT
   5FDI-FFE d
3. 1 ݕࡧઃܭ͔Βਪ࿦ઃܭ΁ͷॏ৺Ҡಈͱ3FDBMM'JSTU
   3FUSJFWBM TODAY'S THESIS RAG ͷ඼࣭վળͷॏ৺͸ɺ ʮݕࡧઃܭʯ͔Β ʮਪ࿦ઃܭʯ΁ҠΓͭͭ͋Δɻ ͦͷస׵ͷഎܠɾख๏ɾͦͯ͠ Recall-First Retrieval ͱ͍͏ઃܭࢥ૝Λ੔ཧ͠·͢ɻ
4. 2 ݕࡧઃܭ͔Βਪ࿦ઃܭ΁ͷॏ৺Ҡಈͱ3FDBMM'JSTU
   3FUSJFWBM
5. 
   ݕࡧઃܭͷຽओԽͱɺͦͷนɻ ݕࡧઃܭ͔Βਪ࿦ઃܭ΁ͷॏ৺Ҡಈͱ3FDBMM'JSTU
   3FUSJFWBM
6. 3 ݕࡧઃܭ͔Βਪ࿦ઃܭ΁ͷॏ৺Ҡಈͱ3FDBMM'JSTU
   3FUSJFWBM RAGͷ
 ࣮૷͸ඪ४Խ͞Εͨɻ CHAPTER 1 · DEMOCRATIZATION
7. 4 ݕࡧઃܭ͔Βਪ࿦ઃܭ΁ͷॏ৺Ҡಈͱ3FDBMM'JSTU
   3FUSJFWBM ͔͠͠ݕࡧͷνϡʔχϯά͸ऴΘΒͳ͍ɻ CHAPTER 1 · THE WALLS
8. 
   ݕࡧઃܭ͔Βɺਪ࿦ઃܭ΁ ݕࡧઃܭ͔Βਪ࿦ઃܭ΁ͷॏ৺Ҡಈͱ3FDBMM'JSTU
   3FUSJFWBM
9. 6 ݕࡧઃܭ͔Βਪ࿦ઃܭ΁ͷॏ৺Ҡಈͱ3FDBMM'JSTU
   3FUSJFWBM • 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
10. 7 ݕࡧઃܭ͔Βਪ࿦ઃܭ΁ͷॏ৺Ҡಈͱ3FDBMM'JSTU
    3FUSJFWBM CHAPTER 3 · CONDITIONS ॏ৺͕ಈ͘3ͭͷ৚݅ɻ
11. 8 ݕࡧઃܭ͔Βਪ࿦ઃܭ΁ͷॏ৺Ҡಈͱ3FDBMM'JSTU
    3FUSJFWBM CHAPTER 3 · The Bitter Lesson RICH SUTTON · 2019 ʮ௕ظతʹ͸ɺ൚༻తͳܭࢉΛ εέʔϧͤ͞Δํ๏͕ɺ ਓ͕ؒखͰ૊ΈࠐΜͩ஌ࣝΛӽ͑Δɻʯ The Bitter Lesson [2]
12. 
    ਪ࿦ઃܭͷͭͷख๏ɻ ݕࡧઃܭ͔Βਪ࿦ઃܭ΁ͷॏ৺Ҡಈͱ3FDBMM'JSTU
    3FUSJFWBM
13. 9 ݕࡧઃܭ͔Βਪ࿦ઃܭ΁ͷॏ৺Ҡಈͱ3FDBMM'JSTU
    3FUSJFWBM
14. 10 ݕࡧઃܭ͔Βਪ࿦ઃܭ΁ͷॏ৺Ҡಈͱ3FDBMM'JSTU
    3FUSJFWBM
15. 11 ݕࡧઃܭ͔Βਪ࿦ઃܭ΁ͷॏ৺Ҡಈͱ3FDBMM'JSTU
    3FUSJFWBM
16. 12 ݕࡧઃܭ͔Βਪ࿦ઃܭ΁ͷॏ৺Ҡಈͱ3FDBMM'JSTU
    3FUSJFWBM • Recall-FirstͰͱ͖ͬͯͨυΩϡϝϯτΛਫ਼ಡ͢ΔͨΊͷίϯϙʔωϯτͱͯ͠ར༻͢Δ ࠷ۙ͸LLM WikiͳͲNavigateϕʔεͷख๏ͷఏҊ͕૿͖͑ͯͨ
17. 
    3FDBMM'JSTU
    3FUSJFWBM ݕࡧઃܭ͔Βਪ࿦ઃܭ΁ͷॏ৺Ҡಈͱ3FDBMM'JSTU
    3FUSJFWBM
18. 13 ݕࡧઃܭ͔Βਪ࿦ઃܭ΁ͷॏ৺Ҡಈͱ3FDBMM'JSTU
    3FUSJFWBM
19. 14 ݕࡧઃܭ͔Βਪ࿦ઃܭ΁ͷॏ৺Ҡಈͱ3FDBMM'JSTU
    3FUSJFWBM TRADE-OFFS ͨͩ͠ɺ ۜͷ஄ؙͰ͸ͳ͍ɻ
20. 1 ݕࡧઃܭ͔Βਪ࿦ઃܭ΁ͷॏ৺Ҡಈͱ3FDBMM'JSTU
    3FUSJFWBM ݕࡧઃܭ͔Βਪ࿦ઃܭ΁ͷॏ৺Ҡಈͱ3FDBMM'JSTU
    3FUSJFWBM
21. 16 ݕࡧઃܭ͔Βਪ࿦ઃܭ΁ͷॏ৺Ҡಈͱ3FDBMM'JSTU
    3FUSJFWBM ͱ͍͏࿩Λٕज़ॻయͰॻ͖·ͨ͠ https://techbookfest.org/product/jbe1zdnjqbb6cHwebRhapn?productVariantID=3VefBkLxtiAHJr1vAHjJMY
22. 17 ݕࡧઃܭ͔Βਪ࿦ઃܭ΁ͷॏ৺Ҡಈͱ3FDBMM'JSTU
    3FUSJFWBM 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
23.  ݕࡧઃܭ͔Β
    ਪ࿦ઃܭ΁ͷॏ৺Ҡಈͱ
    3FDBMM'JSTU
    3FUSJFWBM