Horatio
The last thing you want when you're exploring technical documentation is false information. I'm working on Horatio, an experimental retrieval system for tech doc retrieval that's designed to avoid hallucinations as a first principle. It processes content into a vector store, and a natural language interface uses an LLM to synthesize answers grounded strictly in the retrieved content. It delivers citations to the source pages, assuming it finds some. If it doesn't have the answer, it reports that and mentions anything it found about possible other information sources to check (it doesn't always find those either).
Testing results so far are quite good. Next I plan to add a knowledge graph layer for structured symbolic queries.
The grand plan here is to try to use symbolic AI for what it's good at, and generative AI for what it's good at. And avoid the complex usage issues of the first and the nondeterministic, unreliable output of the second.
This phase focuses on identification and retrieval. If this works out, the next puzzle to embark on is whether generative AI is an effective tool when creating technical documents, and if so, good ways to use it. Some of my assumptions include a rather dim view of current professional tech writing tools. They emphasize the wrong aspects of the documentation process, IMO. Maybe we can do better. And maybe generative AI will be part of that.
But this is an experiment, not product development. The hypothesis is that generative AI can be helpful when writing documents, and it might be disproved.