Slide 1

Slide 1 text

confidencial LLM؂ࢹͷ࠷લઢ IVRy ΤϯδχΞ๨೥LTେձ 2024/12/11 Moriya Hiroyuki

Slide 2

Slide 2 text

confidencial 2 Զͷ໊લ͸ɺߴߍੜ୳ఁ޻౻৽Ұ

Slide 3

Slide 3 text

confidencial 3 Զͷ໊લ͸ɺߴߍੜ୳ఁ޻౻৽Ұ AIΤϯδχΞͱͯ͠ಇ͖࢝ΊͨԶ͸ɺLLMΛ࢖ͬͯɺͨ͘͞Μ͓ۚΛՔ͍Ͱ͍Δε λʔτΞοϓΛ໨ܸͨ͠ɻ

Slide 4

Slide 4 text

confidencial 4 Զͷ໊લ͸ɺߴߍੜ୳ఁ޻౻৽Ұ AIΤϯδχΞͱͯ͠ಇ͖࢝ΊͨԶ͸ɺLLMΛ࢖ͬͯɺͨ͘͞Μ͓ۚΛՔ͍Ͱ͍Δε λʔτΞοϓΛ໨ܸͨ͠ɻ ʮͪΐͪΐͬͱ։ൃͨ͠ΒϘϩṶ͚Ͱ͖Δ΍Μʂʯͱؾ͍ͮͨԶ͸ɺىۀͯ͠ɺͻͨ ͢ΒPoCϓϩδΣΫτΛΫϥΠΞϯτʹఏڙ͢Δ͜ͱʹͨ͠ɻ

Slide 5

Slide 5 text

confidencial 5 Զͷ໊લ͸ɺߴߍੜ୳ఁ޻౻৽Ұ AIΤϯδχΞͱͯ͠ಇ͖࢝ΊͨԶ͸ɺLLMΛ࢖ͬͯɺͨ͘͞Μ͓ۚΛՔ͍Ͱ͍Δε λʔτΞοϓΛ໨ܸͨ͠ɻ ʮͪΐͪΐͬͱ։ൃͨ͠ΒϘϩṶ͚Ͱ͖Δ΍Μʂʯͱؾ͍ͮͨԶ͸ɺىۀͯ͠ɺͻͨ ͢ΒPoCϓϩδΣΫτΛΫϥΠΞϯτʹఏڙ͢Δ͜ͱʹͨ͠ɻ Զ͸ɺഎޙ͔Β͍ۙͮͯ͘Δrate limit੍ݶͱɺLatencyͷѱԽʹؾ͕͍͍ͭͯͳ͔ͬ ͨɻ

Slide 6

Slide 6 text

confidencial 6 Զͷ໊લ͸ɺߴߍੜ୳ఁ޻౻৽Ұ AIΤϯδχΞͱͯ͠ಇ͖࢝ΊͨԶ͸ɺLLMΛ࢖ͬͯɺͨ͘͞Μ͓ۚΛՔ͍Ͱ͍Δε λʔτΞοϓΛ໨ܸͨ͠ɻ ʮͪΐͪΐͬͱ։ൃͨ͠ΒϘϩṶ͚Ͱ͖Δ΍Μʂʯͱؾ͍ͮͨԶ͸ɺىۀͯ͠ɺͻͨ ͢ΒPoCϓϩδΣΫτΛΫϥΠΞϯτʹఏڙ͢Δ͜ͱʹͨ͠ɻ Զ͸ɺഎޙ͔Β͍ۙͮͯ͘Δrate limit੍ݶͱɺLatencyͷѱԽʹؾ͕͍͍ͭͯͳ͔ͬ ͨɻ ؾ͕͍ͭͨΒɺԶͷϓϩμΫτ͸ղ໿͕૬࣍͗ɺձࣾ͸౗࢈ͯ͠͠·͍ͬͯͨ...

Slide 7

Slide 7 text

confidencial 7 ׬

Slide 8

Slide 8 text

confidencial 8 ࠓ೔͸ɺ޻౻৽Ұ܅͕ɺ͜Μͳ݁຤Λܴ͑ͳ͍ͨΊʹͰ͖Δ͜ͱΛ͓࿩͠͠·͢ɻ

Slide 9

Slide 9 text

confidencial ࣗݾ঺հ 2024/08 ೖࣾ SWEɾػցֶशΤϯδχΞͳͲΛܦݧ LLM͕ίΞʹͳΓͦ͏ͳαʔϏεͩͱࢥͬͯIVRyʹೖࣾ Moriya Hiroyuki 9 AI engineer

Slide 10

Slide 10 text

confidencial IVRyͰͷLLMΛར༻ͨ͠AIର࿩ 10 WebsocketΛར༻͠ΤϯυϢʔβʔͱLLM͕ϦΞϧλΠϜʹ΍ΓऔΓ͍ͯ͠Δ

Slide 11

Slide 11 text

confidencial LLM Fallback 11 ෳ਺ͷLLMΛར༻͢Δ͜ͱΛલఏʹFallbackػߏΛߏங APIͷStatus, Ratelimit΍σʔλ੍໿(஍ཧ੍໿)Λ΋ͱʹৼΓ෼͚

Slide 12

Slide 12 text

confidencial LLM Fallback 12 ෳ਺ͷLLMΛར༻͢Δ͜ͱΛલఏʹFallbackػߏΛߏங APIͷStatus, Ratelimit΍σʔλ੍໿(஍ཧ੍໿)Λ΋ͱʹৼΓ෼͚ ؂ࢹ͢Ε͹ ྑ͍ͷ͡Ό

Slide 13

Slide 13 text

confidencial ํ๏ 1ɿDataDog LLM observability 13 DataDog͕Ӷҙ։ൃதͷLLM؂ࢹʹಛԽͨ͠ػೳɻ Latency, token਺, promptͳͲΛऔಘͰ͖Δɻ

Slide 14

Slide 14 text

confidencial 14 ʮ͜ΕͰɺOpenAIͷlatency͕؂ࢹͰ͖ΔΑ͏ʹͳͬͨͥʂʯ

Slide 15

Slide 15 text

confidencial 15 ͋ΕΕʙɺ͓͔͍͠Αʙ

Slide 16

Slide 16 text

confidencial 16 ͋ΕΕʙɺ͓͔͍͠Αʙ ๻ΒͷϓϩμΫτ͸ɺfallbackػߏΛ ࣮૷͍ͯ͠Δͷʹɺ OpenAIͷlatency͔͠؂ࢹͰ͖ͯͳ͍Αʙ

Slide 17

Slide 17 text

confidencial ํ๏ 2ɿOpenLIT (OpenTelemetry) 17 OpenTelemetryن֨ʹଇͬͨɺLLM؂ࢹʹಛԽͨ͠πʔϧɻ ༷ʑͳLLMΛ؂ࢹ͢Δ͜ͱ͕Ͱ͖Δɻ

Slide 18

Slide 18 text

confidencial 18 ʮ͜ΕͰɺ৭ʑͳmodelͷlatency͕؂ࢹͰ͖ΔΑ͏ʹͳͬͨͥʂʯ

Slide 19

Slide 19 text

confidencial 19 ͋ΕΕʙɺ͓͔͍͠Αʙ

Slide 20

Slide 20 text

confidencial 20 ͋ΕΕʙɺ͓͔͍͠Αʙ ๻Β͸ɺ৭ʑͳϞσϧΛ࢖͏Μ͔ͩΒɺ provider͝ͱʹɺlatencyΛܭଌ͢Δඞཁ͕͋Δͷʹ LiteLLMશମͰͷlatency͔͠औΕͯͳ͍Αʙ

Slide 21

Slide 21 text

confidencial ํ๏ 3ɿDataDog Inferred services 21 DataDogʹ౥ࡌ͞ΕͨɺApp֎΁ͷϦΫΤετΛ؂ࢹͯ͘͠ΕΔػߏ

Slide 22

Slide 22 text

confidencial 22 ʮ͜ΕͰɺLiteLLMͰ࢖͍ͬͯΔ͢΂ͯͷmodelΛ؂ࢹͰ͖ΔΑ͏ʹͳͬͨͥʂʯ

Slide 23

Slide 23 text

confidencial 23 ͋ΕΕʙɺ͓͔͍͠Αʙ

Slide 24

Slide 24 text

confidencial 24 ͋ΕΕʙɺ͓͔͍͠Αʙ ๻Β͸ɺGeminiɺOpenAIͰ̍ͭͷmodelΛ ࢖͏ͱ͸ݶΒͳ͍ͷʹɺ ݸผͷmodelͷlatencyΛऔಘ͢Δ͜ͱ͸ Ͱ͖ͯͳ͍Αʙ

Slide 25

Slide 25 text

confidencial ·ͱΊ LLM؂ࢹ͸ɺ·ͩ·ͩൃల్্Ͱ஌ݟ͕͋Γ·ͤΜʂ AIɾLLMΛ࢖͍͜ͳͯ͠ϓϩμΫτʹೖΕ͍ͯ͘աఔͰɺ ࣗΒ͕੾Γ։͍͍ͯ͘ඞཁ͕͋Γ·͢ɻ ͥͻҰॹʹAI؂ࢹΛ΍͍͖ͬͯ·͠ΐ͏ʂ 25