Data ingestion for GPT apps

[opal plaza]

Github repo 👉 https://github.com/ai-sidekick/sidekick

We've been building customer support chatbots using GPT and Weaviate for a few weeks now. We learned that the most time consuming part of making these bots was actually not prompt engineering (which isn't that useful to begin with) but in actually transforming the data into a size and format that works well with GPT's limited prompt window.

We also saw a lot of companies trying to build this out in-house, so we figured we may as well make everything open source. This is mostly some core functionality wrapped around a vector store and modular connectors and parsers. If this sounds useful, check out the repo!

Connectors we've built:

• Github Repo
• HTTP
• Discourse (coming soon)
• Github Issues (coming soon)
• Zendesk (coming soon)

Parsers:

• Markdown
• HTML

We'd love the community's feedback, and of course feel free to fork and use it if it'll be useful for your projects. We tried to make it as easy as possible to add new connectors and parsers.

[nova wyvern]

Hey can I train it on data of my own organisation and then use it to ask questions about ongoing projects?

@opal plaza I am using in context learning and got 3.5 api to make a question and answer bot

But I am facing a problem, how to give it the data of previously asked questions so that it can answer accordingly

[opal plaza]

@nova wyvern yup that's the exact use case that we built sidekick for. We can add a connector for previously asked questions, feel free to make a request in our Github Issues: https://github.com/ai-sidekick/sidekick/issues

[nova wyvern]

@opal plaza thanks for replying can you expalin me how this connector thing works in this case , any docs related to that would be appreciatated

[opal plaza]

We have a docs site coming soon! In the meantime you can ping the API endpoints using an API key directly at https://sidekick-server-ezml2kwdva-uc.a.run.app/docs#/

We currently have 3 endpoints:

/upsert-web-data: This endpoint takes a url as input, uses Playwright to crawl through the webpage (and any linked webpages), and loads them into the vectorstore.

/query: Endpoint to query the vector database with a string. You can filter by source type (web, markdown, etc.) and set the max number of chunks returned.

/ask-llm: Endpoint to get an answer to a question from an LLM, based on the data in the vectorstore. In the response, you get back the sources used in the answer, the user's intent, and whether or not the question is answerable based on the content in your vectorstore.

We can set you up with an API key in our slack support server, you can join that here: https://join.slack.com/t/sidekick-public/shared_invite/zt-1n3iw8via-9y1mP3tJ3~Zy6GT5sZNwOg

[brazen karma]

Good work @opal plaza

[opal plaza]

Thanks @brazen karma !

[nova wyvern]

Thanks

[nova wyvern]

@opal plaza one last doubt bro, I have a model which takes string input and generates output using api, but it runs on vs code terminal, how do i run it on local host or something

been struggling with it

[opal plaza]

@nova wyvern you can use FastAPI to create API endpoints and serve them using uvicorn: https://github.com/ai-sidekick/sidekick/blob/main/sidekick-server/server/main.py#L145

[nova wyvern]

Alright, 👍

[nova wyvern]

@opal plaza I managed to solve my above challange by making a function and storing the previous questions there and then passing them as prompt each time a new question is asked,
Now Can you tell me how to link it to a website

Like how make the whole model available as a callable api

I have account on aws where I can upload the model and index

And for some reason If I ask questions out of context it adds a \n with the answer like this was the response when I asked who is micky mouse:
"\nMicky Mouse is a cartoon character created by Walt Disney in 1928."