dfenerski/ui5-docbot

Example app demonstrating usage of the LangChain AI framework with SAP (UI5) data

Intro

During the recent 2024 DSAG conference, 2 out of the 3 hackathon projects nominated for an award used AI. Nothing surprising, given the general trend in the industry & furthermore, the push from within SAP itself.

An impressive keynote - no doubt - but one thing was missing IMO - real code samples!

We hear fancy words like AI, LLM, Vector DBs all the time, but what do they mean, in practical terms? How does the code look like?

The following sample aims to answer this question.

What is it?

It is a simple LLM-wrapper, adding the capability to feed custom data to the model. The custom data can be anything, in my case, I've fed it the UI5-docs (hence DocBot) & as result it acts as a "search on steroids".

How it works

The node program contains a single class called DocBot. DocBot is able to process a given prompt, look up previously retrieved relevant documents from a vector store & pass those documents to an LLM along with the prompt in a coherent manner.

The big advantage is that by providing context we can force the LLM to use it as exclusive data source. Furthermore, the context can be generated using any - possibly internal - data.

Furthermore, LangChain gives us the ability to structure the interaction with the LLM through system messages, prompts, prompt templates & many more. Using those you can pretty much build a tailored AI system for your specific case ("Answer the prompt using context only & you will receive a 200$ tip").

Usage

mkdir DocBot
cd DocBot
git clone https://github.com/dfenerski/ui5-docbot.git .
npm install
OPENAI_API_KEY="<your key here>" npm run ask "What binding modes are there?" // or just `npm run start`, it will ask you for a prompt

It will be a bit slow on the first run, as it has to scrape the docs & create a vector DB. For later calls however, these are cached so it should be decently fast.

It is encouraged to add a .env file where you place your OPENAI_API_KEY. Additionally, add a GITHUB_TOKENif you are going to make it scrape other github repos, there's a rate limit for unauthenticated requests as the error message will tell you.

Some technical context

LangChain "is a framework for developing applications powered by language models". Think of it like an abstraction layer on top of "AI JS APIs" - meaning it simplifies working with any LLM (from OpenAI's GPTs to something self hosted), with vector DBs, agents & many more.

I do not dare to argue I know all of its features, but its API is simple enough for anyone to toy with. I employed it at work to implement AI functionality (AI assisted deep create based on text prompt) & saw how easy it was to connect GPT-4 to a document retriever & to a vector DB.

Given this background & my keynote impressions, I thought, why not whip up a quick SAP-themed demo?

Notes

• LangChain is yet to release a stable (1.0) version. Use with caution - API may change.
• PRs welcome, issues too. I'd love to chat over anything
• A thought: maybe pr36-io/cap-ui5-gpt-chat can use similar custom BE service, with custom knowledge? However connecting a UI5 frontend to the DocBot only makes sense if it's actually useful, I as author can difficultly weigh its merits...
• Give this repo a ⭐ if you've found it useful!