Skip to main content


Analyzing and improving GPT performance

There are two possible options to improve the performance of your GPT chat:

When a message produced by GPT doesn't meet your expectations, the first thing to do is examine the input that GPT received which led to its response. To find the GPT input, you can go to "Knowledge" -> "Messages". When you hover a message the button "Show prompt" appears. Clicking this button will show you the GPT input text.

Then you can check the sources ("Context") generated from your knowledge database that were used to generate the answer. First you should make sure that these sources all meet your expectations. Otherwise it is impossible for GPT to generate an answer that does. Sometimes it can happen that - for example - out-of-date information is still present in the sources or the wording in the sources is not optimal. In such cases you should first update the sources. To update your sources, you can directly edited any articles you created in LoyJoy. If you are using website sources, you can either exclude specific urls or use CSS selectors to specify which parts of your website should be crawled.

If the sources contain all the required information, but the answer still does not meet your expectations, it might be the case that the prompt and system message need to be optimized. You can take your GPT input from above and enter it in the OpenAI Playground. There you can work on different versions of prompt and system message and immediately see if it generates the expected result.

Questions and answers

How does the prompt and system message for GPT work?

Prompt and system message are used to tell the GPT system how it should generate its answer. By default, we tell the GPT to only refer to the knowledge database ("Context") we feed in to generate its answer. You can also influence the formatting, tone, and length of the answers generated by GPT. You can also tell GPT to e.g. include markdown links in its answer.

Can I activate different articles / sources per process?

Yes, please refere to this page.

How does the crawling work?

Please refere to this page.

How are answers generated?

First LoyJoy searches your knowledge database for the most relevant text snippets. Then these text snippets, your prompt and system message, and the previous question and answer are input to the selected GPT model to generate your answer. See above for further info.

How can I improve answers generated by GPT?

Please refer to this section.