This repository contains all code related to the kickstartDS Concierge hosted at https://www.kickstartDS.com/concierge/
Read more about its release on our blog post:
https://www.kickstartDS.com/blog/launching-the-design-system-concierge/
For a more detailed overview of the architecture, design decisions and some code... have a look at the following Miro board:
https://miro.com/app/board/uXjVMdpB0ao=/
The Concierge can answer user questions by pulling from a big Design System-related knowledge base, gained by scraping manually curated domains on the web.
Crawled HTML gets processed, sections of content are split from it along headline elements, and some metadata is enriched. Finally those sections get transformed to an embedding vector space that is saved to a PostgreSQL database using the pgvector extension.
Requests from the page are sent to a Supabase Edge Function, which get's an embedding for the question by calling the Flask webservice contained within this repository. This embedding is then sent to the database to get back the most relevant sections in relation to the question. This then, finally, gets embedded as context into a GPT Completion prompt, which generates the final answer being streamed back to the user. Additionally links and a preview to all referenced pages gets included as context.
Run app first:
FLASK_APP=app flask run
Now send questions to the locally running service:
curl -d '{"question": "What is a Design System?"}' -H "Content-Type: application/json" -X POST http://127.0.0.1:5000
Install flyctl first. E.g. Arch Linux:
trizen -S flyctl-bin
Log in to your account with flyctl:
flyctl auth login
Deploy a version:
flyctl deploy
Query the running webservice:
curl -d '{"question": "What is a Design System?"}' -H "Content-Type: application/json" -X POST https://question-embedding.fly.dev
This project uses PostgreSQL for data storage, and pgvector as an extension specifically:
https://github.com/pgvector/pgvector
Needed tables are generated by notebooks/create_knowledge_base.ipynb when seed = True is set in cell "Write embeddings to DB".
But you'll need the following pgsql database function:
drop function match_sections (
query_embedding vector(768),
similarity_threshold float,
match_count int
);
create or replace function match_sections (
query_embedding vector(768),
similarity_threshold float,
match_count int
)
returns table (
id bigint,
tokens integer,
content text,
page_url text,
page_title text,
page_summary text,
similarity float
)
language plpgsql
as $$
begin
return query
select
sections.id,
sections.tokens,
sections.content,
sections.page_url,
sections.page_title,
sections.page_summary,
1 - (sections.embedding <=> query_embedding) as similarity
from sections
where 1 - (sections.embedding <=> query_embedding) > similarity_threshold
order by sections.embedding <=> query_embedding
limit match_count;
end;
$$;You will need an account with Supabase to use this repository for yourself directly. But it should be easily adaptable to other hosters, too!
- Go to the "SQL editor" section.
- Click "New Query".
- Enter the above SQL to create or replace your Database function.
- Click "Run" or cmd+enter (ctrl+enter).
See also: https://supabase.com/docs/guides/database/functions
- Install Supabase CLI: https://supabase.com/docs/guides/cli
- Deploy edge function:
yarn supabase functions deploy answer
Optionally integrate the Deno language server with your editor for autocompletion, etc:
https://deno.land/manual/getting_started/setup_your_environment
See also: https://supabase.com/docs/guides/functions/quickstart
yarn supabase gen types typescript --db-url postgres://postgres:YOUR_POSTGRES_PASS@YOUR_POSTGRES_URL:5432/postgres > dbTypes.ts
Unless you explicitly state otherwise, any contribution intentionally submitted for inclusion in the work by you, as defined in the Apache-2.0 license, shall be dual licensed as below, without any additional terms or conditions.
© Copyright 2023 Jonas Ulrich, kickstartDS by ruhmesmeile GmbH [[email protected]].
This project is licensed under either of
at your option.
The SPDX license identifier for this project is MIT OR Apache-2.0.