Building Knowledge, note #1: fixing the gaps
An agent's answer can't be better than the knowledge it has access to. Since this is critical for Ari, Sidekick and every other agent running on Plain, we've started treating knowledge as a product of its own. The goal is a self-improving knowledge loop: it learns from the answers your team has already written, and when an agent can't answer something, you see what's missing, you fill it in, and next time the agent can answer.
The first part we built is Knowledge Gaps. When Ari can't answer, it hands off to a human and records the question. Plain then classifies the question, and if it's a general product question your docs should cover, it creates a knowledge gap. This is necessary as not all unanswered questions are knowledge gaps. For example, "Can you escalate this?" is someone asking for an action and "What plan am I on?" needs access to the customer's account, which Ari doesn't have yet.
Then Sidekick helps you fill it. It reads the threads linked to the gap, past closed threads, your connected docs and any MCP knowledge tools. It then drafts an article with citations and asks where to save it. By default that's your Help Center, but if you've connected another knowledge system over MCP (e.g. Mintlify), you can save it there instead.
Here's a quick demo: https://supercut.ai/share/plain/1MGr7Kz0eOxNyvWfenZrxf
Most of the times, the answer already exists, in a previous thread or in someone’s head. We want to make getting it written down easy and Knowledge Gaps is the first step.