Upgrade to Pro
— share decks privately, control downloads, hide ads and more …
Speaker Deck
Features
Speaker Deck
PRO
Sign in
Sign up for free
Search
Search
IVRyエンジニア忘年LT大会2024 LLM監視の最前線
Search
Hiroyuki Moriya
December 11, 2024
1
240
IVRyエンジニア忘年LT大会2024 LLM監視の最前線
Hiroyuki Moriya
December 11, 2024
Tweet
Share
More Decks by Hiroyuki Moriya
See All by Hiroyuki Moriya
kueueに新しいPriorityClassを足した話
gekko0114
0
660
JobSet超入門
gekko0114
1
850
Featured
See All Featured
Performance Is Good for Brains [We Love Speed 2024]
tammyeverts
7
630
What's in a price? How to price your products and services
michaelherold
244
12k
Done Done
chrislema
182
16k
Adopting Sorbet at Scale
ufuk
74
9.2k
[RailsConf 2023 Opening Keynote] The Magic of Rails
eileencodes
28
9.3k
We Have a Design System, Now What?
morganepeng
51
7.4k
The Success of Rails: Ensuring Growth for the Next 100 Years
eileencodes
44
7k
実際に使うSQLの書き方 徹底解説 / pgcon21j-tutorial
soudai
174
51k
How GitHub (no longer) Works
holman
314
140k
Understanding Cognitive Biases in Performance Measurement
bluesmoon
27
1.6k
Let's Do A Bunch of Simple Stuff to Make Websites Faster
chriscoyier
507
140k
Sharpening the Axe: The Primacy of Toolmaking
bcantrill
40
2k
Transcript
confidencial LLMࢹͷ࠷લઢ IVRy ΤϯδχΞLTେձ 2024/12/11 Moriya Hiroyuki
confidencial 2 Զͷ໊લɺߴߍੜ୳ఁ౻৽Ұ
confidencial 3 Զͷ໊લɺߴߍੜ୳ఁ౻৽Ұ AIΤϯδχΞͱͯ͠ಇ͖࢝ΊͨԶɺLLMΛͬͯɺͨ͘͞Μ͓ۚΛՔ͍Ͱ͍Δε λʔτΞοϓΛܸͨ͠ɻ
confidencial 4 Զͷ໊લɺߴߍੜ୳ఁ౻৽Ұ AIΤϯδχΞͱͯ͠ಇ͖࢝ΊͨԶɺLLMΛͬͯɺͨ͘͞Μ͓ۚΛՔ͍Ͱ͍Δε λʔτΞοϓΛܸͨ͠ɻ ʮͪΐͪΐͬͱ։ൃͨ͠ΒϘϩṶ͚Ͱ͖ΔΜʂʯͱؾ͍ͮͨԶɺىۀͯ͠ɺͻͨ ͢ΒPoCϓϩδΣΫτΛΫϥΠΞϯτʹఏڙ͢Δ͜ͱʹͨ͠ɻ
confidencial 5 Զͷ໊લɺߴߍੜ୳ఁ౻৽Ұ AIΤϯδχΞͱͯ͠ಇ͖࢝ΊͨԶɺLLMΛͬͯɺͨ͘͞Μ͓ۚΛՔ͍Ͱ͍Δε λʔτΞοϓΛܸͨ͠ɻ ʮͪΐͪΐͬͱ։ൃͨ͠ΒϘϩṶ͚Ͱ͖ΔΜʂʯͱؾ͍ͮͨԶɺىۀͯ͠ɺͻͨ ͢ΒPoCϓϩδΣΫτΛΫϥΠΞϯτʹఏڙ͢Δ͜ͱʹͨ͠ɻ Զɺഎޙ͔Β͍ۙͮͯ͘Δrate limit੍ݶͱɺLatencyͷѱԽʹؾ͕͍͍ͭͯͳ͔ͬ ͨɻ
confidencial 6 Զͷ໊લɺߴߍੜ୳ఁ౻৽Ұ AIΤϯδχΞͱͯ͠ಇ͖࢝ΊͨԶɺLLMΛͬͯɺͨ͘͞Μ͓ۚΛՔ͍Ͱ͍Δε λʔτΞοϓΛܸͨ͠ɻ ʮͪΐͪΐͬͱ։ൃͨ͠ΒϘϩṶ͚Ͱ͖ΔΜʂʯͱؾ͍ͮͨԶɺىۀͯ͠ɺͻͨ ͢ΒPoCϓϩδΣΫτΛΫϥΠΞϯτʹఏڙ͢Δ͜ͱʹͨ͠ɻ Զɺഎޙ͔Β͍ۙͮͯ͘Δrate limit੍ݶͱɺLatencyͷѱԽʹؾ͕͍͍ͭͯͳ͔ͬ ͨɻ
ؾ͕͍ͭͨΒɺԶͷϓϩμΫτղ͕૬࣍͗ɺձࣾ࢈ͯ͠͠·͍ͬͯͨ...
confidencial 7
confidencial 8 ࠓɺ౻৽Ұ܅͕ɺ͜Μͳ݁Λܴ͑ͳ͍ͨΊʹͰ͖Δ͜ͱΛ͓͠͠·͢ɻ
confidencial ࣗݾհ 2024/08 ೖࣾ SWEɾػցֶशΤϯδχΞͳͲΛܦݧ LLM͕ίΞʹͳΓͦ͏ͳαʔϏεͩͱࢥͬͯIVRyʹೖࣾ Moriya Hiroyuki 9 AI
engineer
confidencial IVRyͰͷLLMΛར༻ͨ͠AIର 10 WebsocketΛར༻͠ΤϯυϢʔβʔͱLLM͕ϦΞϧλΠϜʹΓऔΓ͍ͯ͠Δ
confidencial LLM Fallback 11 ෳͷLLMΛར༻͢Δ͜ͱΛલఏʹFallbackػߏΛߏங APIͷStatus, Ratelimitσʔλ੍(ཧ੍)ΛͱʹৼΓ͚
confidencial LLM Fallback 12 ෳͷLLMΛར༻͢Δ͜ͱΛલఏʹFallbackػߏΛߏங APIͷStatus, Ratelimitσʔλ੍(ཧ੍)ΛͱʹৼΓ͚ ࢹ͢Ε ྑ͍ͷ͡Ό
confidencial ํ๏ 1ɿDataDog LLM observability 13 DataDog͕Ӷҙ։ൃதͷLLMࢹʹಛԽͨ͠ػೳɻ Latency, token, promptͳͲΛऔಘͰ͖Δɻ
confidencial 14 ʮ͜ΕͰɺOpenAIͷlatency͕ࢹͰ͖ΔΑ͏ʹͳͬͨͥʂʯ
confidencial 15 ͋ΕΕʙɺ͓͔͍͠Αʙ
confidencial 16 ͋ΕΕʙɺ͓͔͍͠Αʙ ΒͷϓϩμΫτɺfallbackػߏΛ ࣮͍ͯ͠Δͷʹɺ OpenAIͷlatency͔͠ࢹͰ͖ͯͳ͍Αʙ
confidencial ํ๏ 2ɿOpenLIT (OpenTelemetry) 17 OpenTelemetryن֨ʹଇͬͨɺLLMࢹʹಛԽͨ͠πʔϧɻ ༷ʑͳLLMΛࢹ͢Δ͜ͱ͕Ͱ͖Δɻ
confidencial 18 ʮ͜ΕͰɺ৭ʑͳmodelͷlatency͕ࢹͰ͖ΔΑ͏ʹͳͬͨͥʂʯ
confidencial 19 ͋ΕΕʙɺ͓͔͍͠Αʙ
confidencial 20 ͋ΕΕʙɺ͓͔͍͠Αʙ Βɺ৭ʑͳϞσϧΛ͏Μ͔ͩΒɺ provider͝ͱʹɺlatencyΛܭଌ͢Δඞཁ͕͋Δͷʹ LiteLLMશମͰͷlatency͔͠औΕͯͳ͍Αʙ
confidencial ํ๏ 3ɿDataDog Inferred services 21 DataDogʹࡌ͞ΕͨɺApp֎ͷϦΫΤετΛࢹͯ͘͠ΕΔػߏ
confidencial 22 ʮ͜ΕͰɺLiteLLMͰ͍ͬͯΔͯ͢ͷmodelΛࢹͰ͖ΔΑ͏ʹͳͬͨͥʂʯ
confidencial 23 ͋ΕΕʙɺ͓͔͍͠Αʙ
confidencial 24 ͋ΕΕʙɺ͓͔͍͠Αʙ ΒɺGeminiɺOpenAIͰ̍ͭͷmodelΛ ͏ͱݶΒͳ͍ͷʹɺ ݸผͷmodelͷlatencyΛऔಘ͢Δ͜ͱ Ͱ͖ͯͳ͍Αʙ
confidencial ·ͱΊ LLMࢹɺ·ͩ·ͩൃల్্Ͱݟ͕͋Γ·ͤΜʂ AIɾLLMΛ͍͜ͳͯ͠ϓϩμΫτʹೖΕ͍ͯ͘աఔͰɺ ࣗΒ͕Γ։͍͍ͯ͘ඞཁ͕͋Γ·͢ɻ ͥͻҰॹʹAIࢹΛ͍͖ͬͯ·͠ΐ͏ʂ 25