LumApps (a Digital Workplace) offers a public documentation web site. It’s a gold mine, and we use it a lot, at Ariella Digital Consulting.

So we have

• dumped all of its content (you can use the lumapps toolbox OR use one of our crawlers)
• converted them to markdown
• split them to chunks (choosing the right size is paramount)
• cleaned them up (you know, that “Request a change on this documentation.” and a few others)
• converted the chunks to embeddings (multiqampnet, to be honest - it’s cheaper than OpenAI and does a great great great job)
• stored all the chunks in a vector database - we love Rust, hence qdrant was a match made in Paradise.

This is all for the information storage part

Whenever a user asks a question to our bot, we

• convert the question to embeddings,
• compute the closest items in qdrant,
• pick the 100 highest results.
• prepare a prompt (that’s developer lingo for “format the question in a way that will force OpenAI to answer without faking the answer and highlight the sources”)
  • (the prompt includes as a context some of the previously selected results)
• send all this to our favourite Generative AI

We render on screen the OpenAI response.