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
LangGraphで始めるマルチエージェントシステム
Search
peisuke
January 30, 2024
Technology
13
3.7k
LangGraphで始めるマルチエージェントシステム
生成AI新年会2024 LT資料
LangGraphで始めるマルチエージェントシステム
peisuke
January 30, 2024
Tweet
Share
More Decks by peisuke
See All by peisuke
AutoGenで作るLLM Agen
peisuke
1
1.2k
Self-RAG: Learning to Retrieve, Generate and Critique through Self-Reflections
peisuke
9
1.4k
Mip-NeRF 360: Unbounded Anti-Aliased Neural Radiance Fields
peisuke
0
7.9k
LangChain Toolsの運用と改善
peisuke
4
2.5k
GNeRF: GAN-based Neural Radiance Field without Posed Camera
peisuke
1
690
TTS Skins: Speaker Conversion via ASR
peisuke
0
360
A Quantum Computational Approach to Correspondence Problems on Point Sets
peisuke
0
650
F0-Consistent Many-to-many Non-parallel Voice Conversion via Conditional Autoencoder
peisuke
0
160
YOLACT real-time instance segmentation
peisuke
1
230
Other Decks in Technology
See All in Technology
OCI コスト管理
ocise
1
110
プロデザ! BY リクルートvol.22_様々なプロダクト経験の中で活きたPdMのスキル
recruitengineers
PRO
3
170
「名前解決」から振り返るAmazon VPC
yuki_ink
0
320
Evolving DevOps Teams and Flexible Organizational Culture
kakehashi
1
130
Namespace, Now and Then
tagomoris
0
170
Oracle Exadata Database Service(Dedicated Infrastructure):サービス概要のご紹介
oracle4engineer
PRO
0
9.3k
Binary Authorizationと友達になろう / Let's be friends with Binary Authorization
iselegant
2
130
RAGHack: Kickoff and RAG 101
pamelafox
0
200
Eventual Detection Engineering
ken5scal
0
390
Estrategias de escalabilidade para projetos web
jessilyneh
1
150
エンジニア向け会社紹介資料
caddi_eng
15
250k
contenteditableと向き合う
kikuchikakeru
2
140
Featured
See All Featured
What's in a price? How to price your products and services
michaelherold
241
11k
How to name files
jennybc
75
97k
BBQ
matthewcrist
83
9.1k
VelocityConf: Rendering Performance Case Studies
addyosmani
321
23k
ReactJS: Keep Simple. Everything can be a component!
pedronauck
662
120k
Optimising Largest Contentful Paint
csswizardry
28
2.7k
Design and Strategy: How to Deal with People Who Don’t "Get" Design
morganepeng
122
18k
[RailsConf 2023 Opening Keynote] The Magic of Rails
eileencodes
27
8.9k
Done Done
chrislema
180
16k
Reflections from 52 weeks, 52 projects
jeffersonlam
346
20k
Creatively Recalculating Your Daily Design Routine
revolveconf
215
12k
Put a Button on it: Removing Barriers to Going Fast.
kastner
58
3.4k
Transcript
-BOH(SBQIͰ࢝ΊΔϚϧ νΤʔδΣϯτγεςϜ !QFJTVLF
None
-BOH$IBJOॲཧ͕γʔέϯγϟϧͱ͍͏՝͕͋Δ • $IBJOͰɺ"#$ͱ͍ͬͨॲཧΛߦ͏ • ྫɿݕࡧͯ͠ཁͯ͠༁͢ΔͳͲ • --.ͷΑ͏ʹᐆດͳೖग़ྗΛѻ͏έʔεͰɺదͳ݁Ռ͕ಘ ΒΕͳ͍͜ͱ͕ଟ͍ • ྫɿʮ౦ژͷҎ߱ͷਓޱͷਪҠΛग़ྗͯ͠ʯʹର͠ɺதؒ
ͷॲཧͰҎ߱ͷਪҠΛޡͬͯग़ͯ͠͠·͏ͱमਖ਼͕Ͱ͖ͳ͍
-BOH(SBQI • άϥϑͱͯ͠Ұ࿈ͷಈ࡞Λهड़͠ɺʮαΠΫϧʯͷॲཧΛѻ͑ ΔΑ͏ʹͨ͠ • ҰݸͷॲཧΛʮϊʔυʯͰද͠ɺ࣍ʹͲͷϊʔυΛಈ͔͔͢Λ ʮΤοδʯͰද͢ ௨ৗ'VODUJPO$BMMJOH
-BOH(SBQIͱϚϧνΤʔδΣϯτ • ෳͷϊʔυΛར༻͢Δ͜ͱͰϚϧνΤʔδΣϯτͷγεςϜ ͱͯ͠ར༻Ͱ͖Δ ৭ΜͳߏΛ࡞ΕΔ
ঢ়ଶʢεςʔτʣͱ • άϥϑશମͷใʢաڈͷձཤྺɺೖྗͳͲʣ • جຊతʹࣗ༝ʹهड़ͯ͠ྑ͍ • ޙʹઆ໌͢ΔϊʔυΤοδ͕ར༻͢ΔใΛೖΕ͓ͯ͘
ϊʔυͱ • Կ͔͠ΒͷॲཧΛද͢ʢྫɿ--.ɺ'VODUJPO$BMMJOHʣ • ؔͱͯ͠ఆ͓ٛͯ͘͠ • εςʔτΛೖྗ͠ɺॲཧ݁ՌΛฦ͢
Τοδͱ • ݺͼग़͠ݩͱݺͼग़͠ઌͷϊʔυΛهड़ • ݅ذ͕ඞཁͰ͋Εɺ݅ذͷ໊ؔهड़ TIPVME@DPOUJOVFؔͰ εςʔτΛೖྗͱͯ͠ɺ DPOUJOVF͔FOEΛฦ͢
Τοδͱ • ۩ମతͳ݅ذͷྫ ࠷৽ͷϝοηʔδ͕ 'VODUJPO$BMMJOHΛݺ ΜͰ͍Δ͔ΛνΣοΫ
ಈ࡞ͷΠϝʔδΛղઆ • ఱؾΛௐΔΈΛఆ • "HFOUϊʔυͱɺ5PPMϊʔυΛ༻ҙ • "HFOUɺ5PPMΛݺͿ͔Ͳ͏͔ΛܾΊɺݺ ͳ͍߹ࣗવݴޠͰฦ • 5PPMɺఱؾ"1*Λݺͼग़ͯ͠ճ
"HFOU 5PPM &OE 4UBSU ఱؾΛௐΔ'$
ಈ࡞ͷΠϝʔδΛղઆ • ఱؾΛௐΔΈΛఆ • "HFOUϊʔυͱɺ5PPMϊʔυΛ༻ҙ • "HFOUɺ5PPMΛݺͿ͔Ͳ͏͔ΛܾΊɺݺ ͳ͍߹ࣗવݴޠͰฦ • 5PPMɺఱؾ"1*Λݺͼग़ͯ͠ճ
"HFOU 5PPM &OE 4UBSU ఱؾΛௐΔ'$ ユーザー:東京の天気は? Agent:{Function calling: 天気API, query: 東京}
ಈ࡞ͷΠϝʔδΛղઆ • ఱؾΛௐΔΈΛఆ • "HFOUϊʔυͱɺ5PPMϊʔυΛ༻ҙ • "HFOUɺ5PPMΛݺͿ͔Ͳ͏͔ΛܾΊɺݺ ͳ͍߹ࣗવݴޠͰฦ • 5PPMɺఱؾ"1*Λݺͼग़ͯ͠ճ
"HFOU 5PPM &OE 4UBSU ఱؾΛௐΔ'$ Agent:{Function calling: 天気API, query: 東京} EdgeはAgentの返答がFunction Callingかを判定、 Function Callingが呼ばれたので、Toolノードに遷移
ಈ࡞ͷΠϝʔδΛղઆ • ఱؾΛௐΔΈΛఆ • "HFOUϊʔυͱɺ5PPMϊʔυΛ༻ҙ • "HFOUɺ5PPMΛݺͿ͔Ͳ͏͔ΛܾΊɺݺ ͳ͍߹ࣗવݴޠͰฦ • 5PPMɺఱؾ"1*Λݺͼग़ͯ͠ճ
"HFOU 5PPM &OE 4UBSU ఱؾΛௐΔ'$ Agent:{Function calling: 天気API, query: 東京} Tool: {message: 晴れ} πʔϧݺͼग़͠
ಈ࡞ͷΠϝʔδΛղઆ • ఱؾΛௐΔΈΛఆ • "HFOUϊʔυͱɺ5PPMϊʔυΛ༻ҙ • "HFOUɺ5PPMΛݺͿ͔Ͳ͏͔ΛܾΊɺݺ ͳ͍߹ࣗવݴޠͰฦ • 5PPMɺఱؾ"1*Λݺͼग़ͯ͠ճ
"HFOU 5PPM &OE 4UBSU ఱؾΛௐΔ'$ ユーザー:東京の天気は? Agent:{Function calling: 天気API, query: 東京} Tool: {message: 晴れ} Agent: {message: 東京の天気は晴れ}
ಈ࡞ͷΠϝʔδΛղઆ • ఱؾΛௐΔΈΛఆ • "HFOUϊʔυͱɺ5PPMϊʔυΛ༻ҙ • "HFOUɺ5PPMΛݺͿ͔Ͳ͏͔ΛܾΊɺݺ ͳ͍߹ࣗવݴޠͰฦ • 5PPMɺఱؾ"1*Λݺͼग़ͯ͠ճ
"HFOU 5PPM &OE 4UBSU ఱؾΛௐΔ'$ Agent: {message: 東京の天気は晴れ} Agentの返答にFunction Callingが含まれな いので、END側のノードを呼び出す
ಈ࡞ͷΠϝʔδΛղઆ • Ϧαʔνͯ͠υΩϡϝϯτΛ࡞ΔΈΛͪΐͬͱ͚ͩ
ಈ࡞ͷΠϝʔδΛղઆ • Ϧαʔνͯ͠υΩϡϝϯτΛ࡞ΔΈΛͪΐͬͱ͚ͩ 全体統括者
ಈ࡞ͷΠϝʔδΛղઆ • ΤοδͷΈ͚ͩ؆୯ʹհ リサーチチームの結果は統括者に返す、 ドキュメントチームの結果も統括者に返す 統括者は次のチームをJSONを出⼒する 仕様、レスポンスのnextに応じて、次 のチームを決めてノードを動かす
ಈ࡞ͷΠϝʔδΛղઆ • Ϧαʔνͯ͠υΩϡϝϯτΛ࡞ΔΈΛͪΐͬͱ͚ͩ リサーチチーム ドキュメントチーム 全体統括者
ಈ࡞ͷΠϝʔδΛղઆ • ΤοδͷΈ͚ͩ؆୯ʹհ 先程と同様に、ノードの 結果を統括者に返す 先程と同様に、Search, Web Scraperのノードを 選択的に呼び出す
"VUP(FOͱͷҧ͍ • ݁ہͷͱ͜Ζશ෦ॻ͚ಉ͜͡ͱ͕Ͱ͖ΔΑ • ʮࣗવʹ͑ʯͱ͍͏લఏͰݸਓͷײΛड़Δ • ࡉ͔͘ॻ͘ɺҙͷॲཧΛ࣮ߦ͢Δͷ͕ۤखͦ͏ • ಛʹ"VUP(FOΑΓίʔυΛॻ͍࣮ͯߦ͢Δͷ͕PQUJPOBMѻ͍ •
ભҠΛ໌֬Խ͢Δ͜ͱͰɺ੍ޚ͕༰қʹͳΔ • ͋Δఔɺܾ·ͬͨۀϑϩʔͷࣗಈԽͰޮՌΛൃشͦ͠͏
·ͱΊ • -BOH(SBQIͷجຊతͳߏͱɺϚϧνΤʔδΣϯτͷࡍͷߏ ʹ͍ͭͯ؆୯ʹհ • ॲཧͷखॱΛৄࡉʹղઆɺΠϝʔδ͕͍ͨΒ͍Ͱ͢