Overview

You can build a self-improving Fibi workflow that turns recurring support questions, corrected answers, and product changes into better knowledge for your AI agent.

The Featurebase MCP server and Training Data API let your AI tool create and update Q&A snippets and training files. Before saving a proposed answer, it can ask the oracle to check existing knowledge and suggest where the information belongs.

The loop is simple: find a gap → propose an answer → check existing knowledge → review and save.

What can you use it for?

• Learn from resolved conversations – Turn reviewed solutions to recurring questions into reusable Q&A snippets.

• Keep knowledge current after releases – Have a coding agent compare product changes with existing knowledge and propose updates.

• Maintain a large Q&A collection – Import approved answers from a spreadsheet or another system, then update the same entries when the source changes.

• Reduce duplicate and conflicting content – Check whether a topic already has an answer before adding another snippet.

For a one-off correction directly in the Inbox, use Improve Fibi's AI answers.

How the oracle keeps knowledge organized

The oracle is a read-only check for a proposed question and answer. It looks for related knowledge, compares selected matches, and returns a suggested next step with a reason.

• Create a new Q&A – No overlap was found among the content checked.

• Update an existing Q&A – The information belongs with an entry that already covers the topic.

• Review a Q&A, article, or file – Related content may conflict with the proposal or need a closer look. Your agent can open the source and suggest an edit in the right place.

• Skip the addition – The information appears to be covered already, or the proposed entry is unchanged.

For example, if a reviewed product change updates your refund policy, the oracle may point to an existing Q&A or Help Center article that still describes the old policy. Your team can update that source instead of adding another answer alongside it.

Setting up your first workflow

1. Connect your AI tool

Follow the MCP setup guide to connect your workspace. Allow the tools needed to read conversations and knowledge, and choose which write actions require confirmation.

If you want your agent to use GitHub, Linear, or another source, it also needs access to that source through your AI tool.

2. Start with a small set of sources

Choose a few resolved conversations, a reviewed release, or an approved Q&A document. Ask your agent to identify reusable facts and show the evidence behind each proposed answer.

Keep customer-specific details, personal information, and secrets out of reusable training content. A solution for one customer may not be a policy that applies to everyone.

3. Check proposals with the oracle

Ask your agent to run each proposed Q&A through the oracle and show you the suggested action, reason, and any existing content it found. Resolve conflicting or uncertain information against your authoritative source before saving.

4. Save approved changes and test

After review, your agent can create or update the relevant Q&A, training file, or Help Center article using its available tools. Q&A snippets and files are managed under Dashboard → Automations → Training Data.

Wait for processing and indexing to complete, then test Fibi with representative questions. Keep a record of the previous content so you can restore it if needed.

Once the process works well, you can schedule it in an automation-capable AI tool. Keep review enabled for changes you want to approve individually.

Example prompts

Find knowledge gaps in support conversations:

Review 10 resolved support conversations from the last week. Find recurring questions where our knowledge is missing or outdated. Propose concise Q&A updates with supporting sources, remove customer-specific details, and check each proposal with the training-data oracle. Show me the suggested action and any conflicting content. Wait for my approval before saving anything.

Keep Fibi in sync with a release:

Review these approved release notes against our existing Fibi knowledge and Help Center articles. Propose the necessary updates, use the oracle for proposed Q&A changes, and show me where each change belongs. Prefer updating the existing source when appropriate. Wait for approval before editing or publishing.

FAQs