| I’ve been tinkering with an LLM-based assistant for small-scale web development team's knowledge base. Picking a tool and setting up a knowledge base to have technical and design decisions, notes, and maintenance logs in a single place is usually straightforward. It's the consistently feeding the system with high-quality data part that we get stuck at. Eventually, I have to input information when I lack the necessary mental bandwidth, I end up creating incomplete entries that lack essential details. The initial idea is to develop a chat interface that allows us to share random pieces of information, either through text or voice, and have it scanned for actionable data that can be ingested. Then it should start probing the user, looking to expand and refine the new information, connecting it with existing data, and general maintenance of the repository. All with no active guidance. For instance, I could launch it and say: “Today the team and I discussed how frequently the databases for project A and B should be synced. We established the user can tolerate up to 10 minutes of tardiness", at which point I would expect it to: - Create a note of this discussion happening
- Check and fix the existence of entries for projects A and B
- Link the note to the projects
- Create database entries for both projects
- Look for previous entries that could be linked to the projects
- Ask for further explanation of how we came to that conclusion |