Slide 1

Slide 1 text

LangChain Japan MeetUp @Mercari, Inc., Roppongi 2023/07/20 Hiroki YAMAMOTO New Bubiness Dep., Classmethod, Inc.

Slide 2

Slide 2 text

2 About me Hiroki YAMAMOTO @Classmethod, Inc. May 2020~ ・R&D team of Computer Vision ・Skeleton Detection ・Person Tracking March 2023~ ・researching LLM My blog posts ・https://dev.classmethod.jp/author/yamamoto-hiroki/

Slide 3

Slide 3 text

3 About Classmethod, Inc. (クラスメソッド株式会社) Main business ・AWS (SI Partners of the Year 2022) Company blog ・DeveloperIO (over 40,000 posts) Focusing on LLM ・OpenAI 632 posts ・ChatGPT 349 posts (as of July 17) Hiring! ・https://careers.classmethod.jp/requirements/nb-llm-engineer/ start consulting service press release (March 16 2023) service for multi GAIs and secure connection press release (July 7 2023)

Slide 4

Slide 4 text

4 We are utilizing LangChain (Thank you!) Session presentations @ devio2023 (more than 500 attendees) Also using on the back-end side De-Blackboxing! Thorough explanation of LangChain and LlamaIndex used with LLM. Extending ChatGPT‘s functionality with LangChain ~How to generate natural responses by leveraging past conversations and Google search results~

Slide 5

Slide 5 text

5 My Interest Agent・Tool(Function Calling) ・I tried with Zapier NLA (March 17 2023) → Awesome AI can have arms and legs! Zapier Natural Language Actions + LangChain is amazing, so I tried it decompose task and call tools

Slide 6

Slide 6 text

6 My Job (I’m researching) QA Chatbot ・RAG (Retrieval Augumented Generation) = LangChain Indexes I created a bot that uses OpenAI's API to perform vector searches of sales materials I tried "except" and "large and small numbers" in the vector search and it looked difficult I've been trying to figure out the problem and how to improve it when a vector search doesn't give me the information I want

Slide 7

Slide 7 text

7 How to improve

Slide 8

Slide 8 text

8 I’m trying Function Calling ・(via LangChain Agent) Experiment ・about ・Numerical Comparison ・Tool Selection ・perspective ・Cost・Performance ・GPT3.5 vs GPT4 I checked the conditions under which a function is called in Function calling of OpenAI API

Slide 9

Slide 9 text

9 My Opinion (I'm searching for / finding / looking forward to) new type of LLM specialized for Agent (will come / change the world) ・very smart (logical, comprehensive, thoughtful) ・not need so much knowledge ・low cost, low latency (for iterative thinking = API call) ・can control tens of tools hierarchical structured tools/agents for more accurate task processing ・like human organizational structure ・each tool may include agent (and human as a tool)