Confluence Optimization Best Pratices

mjohnson

We use confluence as our primary documentation tool and the ability for Glean to search it has been great!

However, I ran into an issue today that I was wondering about how to optimize the results. The context is that I build data pipelines, and I want other employees to be able to understand those processes so they know how the data they are looking at got created.

I create confluence documents to keep notes to myself about potential solutions AND pages that represent the actual in production solutions.

When I asked Glean about a particular data pipeline it returned both my potential solutions and the actual in production solutions which results in an answer that is less useful then it could be if it only focused on the production data.

How can I improve my confluence usage such that Gleans answers will return the data thats actually useful to the question? I'm not sure the answer is truly to exclude pages, but that's conceptually close to what I'm looking for.

Trevor

This is a great question!

There are a few possible actions to take. The choice of action(s) to take will be a effort vs. value tradeoff. The list below is nominally in order of least effort to most effort.

• Use Glean's Verification feature to verify the documents that represent the production solutions. This will boost the production solutions in the search results, which increases the probability that Glean Assistant will use those documents first when responding to user prompts.
• In your Confluence pages, if you're not already doing so, include statements about the purpose of the document. For example, on the documents that represent the potential solutions, be clear about the fact that the documents record potential solutions. In some cases, you may have already organized your Confluence Space to make this clear to a human reader, but that organizational structure may not be represented in Glean's index, so the information isn't available to the LLM to use that distinction when generating its response.
• On the user side, provide guidance about how to prompt Glean Assistant to get the desired response. The user could include language in their prompt like, "Ignore documents that present potential solutions. Only generate the response based on documents that present production solutions." Similarly, let users know that if they get a response that includes documents that it shouldn't, Glean Assistant has a feature where they can remove those documents and regenerate the response.
• Build a chat-based Glean Agent that adds instructions explicitly, so the user doesn't need to. There are many possible advantages to this approach, which could include restricting the data corpus, implementing branching strategies based on the user's prompt, and many more. This approach takes the most effort, creativity, and testing, but if it is a common use case, then optimizing how it works can have a significant ROI.

If you try out any of these ideas, please let us know how well they worked, or if they didn't help for your use case or your enterprise environment.