RAGの精度が全然上がらない!! AOSSを使った社内RAG開発の反省

Slide 1

Slide 1 text

RAGͷਫ਼౓͕શવ্͕Βͳ͍!! AOSSΛ࢖ͬͨࣾ಺RAG։ൃͷ൓ল 2024.12.13 JAWS-UG ਆށ #3 ೥๨ΕେLTେձ @tosuri13

Slide 2

Slide 2 text

ͱ͢Γ @tosuri13 MOTEXגࣜձࣾ ࡶ༻ܥΤϯδχΞ(ࣗশ) Stor a ge Browser for Am a zon S3Λ Amplifyൈ͖Ͱ࢖͑ͳ͍͔໛ࡧதͰ͢🥺

Slide 3

Slide 3 text

ࣾ಺จॻΛѻ͑ΔRAGγεςϜΛAWSͰ։ൃɾӡ༻த… RAGͷਫ਼౓޲্ʹେۤઓ!! ՝୊΍൓লϙΠϯτʹ͍ͭͯ࿩͍ͨ͠ͱࢥ͍·͢ ↑ ࠓ೥7݄ͷBedrock Night in େࡕͰ΋࿩ͨ͠RAGγεςϜͰ͢☺

Slide 4

Slide 4 text

ͬ͘͟Γݱঢ়ͷAWSߏ੒ਤ

Slide 5

Slide 5 text

RAGͷਫ਼౓͕શવ্͕Βͳ͍!!

Slide 6

Slide 6 text

AOSSʹυΩϡϝϯτΛେྔ౤ೖͯ͠ӡ༻։࢝!! → ͔͠͠ظ଴͢Δճ౴͕શવฦͬͯ͜ͳ͍!! ݪҼΛ୳Δ͜ͱʹ… ͜ͷػೳ͕௥Ճ͞Εͨͷͬͯ Ͳͷόʔδϣϯ͔Β? ͹ͳͳ ஌Βͳ͍Ͱ͢ ࣾ಺RAGϘοτ

Slide 7

Slide 7 text

RAGʹ͓͍ͯҰ൪ॏཁͳϑΣʔζ͸Ͳ͔͜? https:// a ws. a m a zon.com/jp/blogs/news/ a -pr a ctic a l-guide-to-improve-r a g-systems-with- a dv a nced-r a g-on- a ws ΑΓҾ༻

Slide 8

Slide 8 text

https:// a ws. a m a zon.com/jp/blogs/news/ a -pr a ctic a l-guide-to-improve-r a g-systems-with- a dv a nced-r a g-on- a ws ΑΓҾ༻ Ұൠతʹ͸RetrieveॲཧͩͱݴΘΕ͍ͯΔ (ແؔ܎ͳσʔλΛ౉ͯ͠͠·͏ͱɺͲΕ͚ͩ׬ᘳͳLLMͰ΋ద੾ͳճ౴͕Ͱ͖ͳ͍ͨΊ)

Slide 9

Slide 9 text

ͦͷͨΊɺRetrieveॲཧΛ࠷దԽ͢ΔΑ͏ʹ ༷ʑͳΞϓϩʔνͰ࣮ݧɺௐࠪΛॏͶΔ… Am a zon BedrockͷຒΊࠐΈϞσϧΛม͑ͯΈͨΓ… (EmbeddingҎ֎ʹ΋ϝλσʔλΛೖΕͨΓ) AOSSͷANNϥΠϒϥϦΛม͑ͨΓ HNSWͷύϥϝʔλΛ͍͔͍ͭͬͯ͘͡ΈͨΓ… F a iss͔nmslib͔…

Slide 10

Slide 10 text

Slide 11

Slide 11 text

ճ౴ͷ඼࣭௿ԼΛট͍͍ͯͨຊ౰ͷݪҼ͸… https:// a ws. a m a zon.com/jp/blogs/news/ a -pr a ctic a l-guide-to-improve-r a g-systems-with- a dv a nced-r a g-on- a ws ΑΓҾ༻

Slide 12

Slide 12 text

ͩͬͨ͜͜!! https:// a ws. a m a zon.com/jp/blogs/news/ a -pr a ctic a l-guide-to-improve-r a g-systems-with- a dv a nced-r a g-on- a ws ΑΓҾ༻

Slide 13

Slide 13 text

ݪҼௐࠪΛਐΊΔ্Ͱ෼͔ͬͨͷ͸… ʮͦ΋ͦ΋ͷυΩϡϝϯτ͕͓͔͍͠!!ʯͱ͍͏͜ͱ ֦ுࢠͱ࣮ࡍͷத਎͕ Կނ͔ҧ͏ϑΝΠϧ ΫΫΫ… ↑ mdʹٖଶ͢Δxml ωਃExcelͷଘࡏ (ҳ୤ͨ͠࢖͍ํͷExcelγʔτ) ແҙຯͳ৘ใͰ ຒΊਚ͘͞ΕͨHTML ϦϯΫେྔʹ࣋ͬͯ·͢

Slide 14

Slide 14 text

ݪҼௐࠪΛਐΊΔ্Ͱ෼͔ͬͨͷ͸… ʮͦ΋ͦ΋ͷυΩϡϝϯτ͕͓͔͍͠!!ʯͱ͍͏͜ͱ ֦ுࢠͱ࣮ࡍͷத਎͕ Կނ͔ҧ͏ϑΝΠϧ ΫΫΫ… ↑ mdʹٖଶ͢Δxml ωਃExcelͷଘࡏ (ҳ୤ͨ͠࢖͍ํͷExcelγʔτ) ແҙຯͳ৘ใͰ ຒΊਚ͘͞ΕͨHTML ϦϯΫେྔʹ࣋ͬͯ·͢ ݕࡧͷϊΠζʹͳΓɺద੾ͳσʔλΛ ఏڙͰ͖ͳ͘ͳ͍ͬͯΔ!!

Slide 15

Slide 15 text

લॲཧ + νϟϯΩϯάΛ؁͘ݟ͍ͯͨͷ͕ݪҼ!! (L a ngCh a inඋ͑෇͚ͷϩʔμʔʹॲཧΛؙ೚͍ͤͯͨ͠΋ѱ͔ͬͨ…) ݕূ࣌ ࣮ӡ༻࣌ ͋Δఔ౓៉ྷͳυΩϡϝϯτΛ ࢖ͬͯݕূͨͨ͠Ί໰୊ʹؾ෇͚ͣ… શવେৎ෉ͩͳ!! ࣮ࡍͷυΩϡϝϯτ͸ۄੴࠞަঢ়ଶ!! اۀͷ࣮ଶʹԊͬͨॲཧΛ૊Ήඞཁ͕͋Δ

Slide 16

Slide 16 text

֤υΩϡϝϯτͱਅ݋ʹ޲͖߹͍ ͦΕͧΕ࠷దͳܗͰม׵͍ͯ͘͠ॲཧΛߦͳͬͨ ࣄલʹυΩϡϝϯτΛ֬ೝ͠ ѻ͑ͳ͍΋ͷ͸ೖΕͳ͍ த਎XML͡ΌΜ!! ωਃ͔Β࢖͑Δ෦෼Λநग़ LLMͰཁ໿ͯ͠Ϩίʔυܗࣜʹ BS4Ͱղੳͯ͠ ෆཁͳλάΛΫϦʔχϯά

Slide 17

Slide 17 text

͢Δͱɺಛʹߏ੒΍ઃఆ͸มΘ͍ͬͯͳ͍ͷʹ ظ଴͢Δճ౴͕໯͑ΔΑ͏ʹͳͬͨ!! ͜ͷػೳ͕௥Ճ͞Εͨͷͬͯ Ͳͷόʔδϣϯ͔Β? όʔδϣϯX.X͔ΒͰ͢ ࣾ಺RAGϘοτ ͍͍ײ͡ͷػೳ͕૿͑·ͨ͠

Slide 18

Slide 18 text

No content

Slide 19

Slide 19 text

·ͱΊ

Slide 20

Slide 20 text

ɾAIٕज़ཁૉʹਅ݋ʹ޲͖߹͓͏!! ɾRAGΛݕ౼͢ΔલʹυΩϡϝϯτ੔ཧΛ!! → ͦ΋ͦ΋RAGΛ࢖Θͣͱ΋ɺ͙͢ʹυΩϡϝϯτΛݟ͚ͭΒΕΔঢ়ଶ͕๬·͍͠Ͱ͢ → ීஈ͔Β៉ྷͳυΩϡϝϯτΛॻ͖·͠ΐ͏!! → લॲཧΛͤͣʹద౰ʹಥͬࠐΉͱμϝͩͱ͍͏ͷ͕Α͘෼͔Γ·ͨ͠ → طଘͷAIαʔϏε೚ͤͰ͸ͳ͘ɺటषͯ͘΋͍͍ͷͰਅ݋ʹ޲͖߹͏͜ͱ͕େࣄͰͨ͠

Slide 21

Slide 21 text

Th a nk you for listening!! @tosuri13 ← Α͔ͬͨΒTwitterϑΥϩʔͯ͠Ͷ