Common Issues & Errors

Creating prompts with AI definitely has a knack to it and you may not find it super simple right away. It is however very rewarding and in a way, similar to writing code. Any single mistake in your prompt could cause an error and these are what we want to avoid. Here we want to provide some of the common issues we encounter and how we should avoid these.

  • Maximum Prompts - Every model provider has a maximum prompt output. When counting the output, they take the prompt into account, this means that you may have 2048 tokens to play with. If your prompt is 1500 tokens and you are trying to generate 700 tokens, you will be over that 2048 total so it is a balancing act of proviidng enough context within the prompt but also allowing enough tokens to be used for a quality output.

A token is roughly 4 characters. What we like to do is to take our prompts and put them into a character counter. Divide the number by 4 and you will have a rough estimate for the token amount. This won't be perfect but will be near enough for you to use when calculating your models and ensuring you don't go over the maximum.

  • Lack of Context - It is easy to think of this technology as groundbreaking and to big it up more than it is. You might make the big brain move of thinking let's just make it write us a full blog and putting in a prompt like 'Write me a full blog on Financial Systems' and then setting the token amount to about 10 less than maximum to take into account that title and hitting generate. The AI will generate those tokens but it has no context and limited idea for what you want so don't expect anything useable to come out. Examples are often required and encouraged for quality prompt building.

  • Filters - Due to the way we have build Riku, there are no filters on the majority of our models. Note that if you are to go to production with anything from OpenAI, you will be required to put a filter on the model but in RIku, we don't use them as we just provide the playground for building out your prompts. Cohere is the only technology that has the filter built directly into the generate endpoint so you may encounter filters with Cohere whilst using it in Riku.

  • API Keys - Check that your API Keys are correct when entering them into Riku. If they are not added correctly, the technology will not work. You can remove the current settings and add your new settings at any time.

Last updated