Ask Liberty AI - Persona variation ideas

Liberty Converse
1

We are currently looking into using the Ask Liberty AI function within Agent Assistant to help our users access our large collection of guidance and reference documents. Our testing with this has been going well and it has shown itself to be very good at searching through large bodies of text for the most relevant information.

A small sticking point for us is the Persona section of the form. We have been able to use this to make small stylistic changes, from bolding email addresses and summarising in bullet points, to giving replies like a pirate (testing environment only!). However, some of our additions have been inconsistently applied, or not at all.

Does anyone have experience of using the Persona within Agent Assistant to set the formatting of responses to agent queries based on Collection PDFs? Are there any things particularly suitable for amending through the portal, and are there any hard rules on what can’t be amended/asked for? Are there any important rules for how we should provide data that can be most easily read by Liberty AI without needing tweaking in the Persona? (We are already working on removing images and simplifying tables.)

Thank you for reading!

2

Hi Robert. Would you be able to provide some example personas that you’ve been using - both ones that are working and ones that are inconsistent. We can then take a look and may be able to advise on some improvements.

In terms of PDFs, the cleaner the text the better. Removing images is a good idea - although note that we hope to support images (and other document types) in future releases. One other important factor is to try to minimise conflicting or overlapping information in the uploaded documents.

3

Hi Sam, thank you for your response. This is an example of a Persona type we have been using:

”You are a supportive and polite bot assisting council staff with technical queries. Keep language simple. Summarise information from the source in complete sentences or bullet points.

When you find an embedded link, extract the link and include this in your response as a link. Add http:// before www. This link should be a clickable URL and should take the user to an external site.

Put all email addresses in bold. Make any URLs clickable links. Remove spaces within a URL.

If the source has indented bullet points, indent these in the response as well.

If a postal address appears on multiple lines, keep the line breaks.”

Our source material contained a lot of embedded links, which we found weren’t being translated into the datasets used by the AI. The first bolded section above was our attempt to work around this issue, though I suspect this may not be possible. Similarly, we had indented bullet points in our guides that do not feed through to the output, but this line of the persona was added before it was noted that indented bullet points weren’t supported within AA and VA.

The italicised section refers to an issue we have with URLs in plain text in the source material PDFs. These are converted into clickable URLs in the output, but if there is a line break in the source, spaces are added or characters removed that render the link invalid. So far, the only consistent workaround we have found to this is to put website URLs in a separate PDF, orientated in Landscape and with the font shrunken enough to fit the address in a single line.

Another concern we have is character limitations on outputs, and how to regulate these. Our source material (which we do plan to review extensively!) has several very long lists of bullet points and multi-step processes. Occasionally we see outputs ending in ‘’, suggesting we have broken something somewhere, possibly by trying to invoke too much in a response. Knowing what the limits are and what we could add to mitigate this would be helpful.

4

Hi Robert,

Thanks for the detailed response; this is really useful, and it’s clear you’ve put a lot of thought into how you’re using the Persona and structuring your documents.

Some of the behaviours you’re seeing (like how links or bullet points are handled) are known limitations or areas where results can vary depending on the source formatting and how the AI interprets it. Your workarounds - like isolating URLs or simplifying formatting - are along the right lines, and we’re continuing to explore ways to make this more consistent in future updates.

Rather than speculate too much here, would you be open to a short session with someone from our AI team? We’d love to walk through your examples, share some tips based on what we’ve seen work well elsewhere, and see how we can help you get the most out of the tool.

I’ll reach out by email to arrange.

Thanks
Sam.

1 Like