Prompt Engineering

With the rise of generative AI models, comes a need to structure text that can be interpreted and understood by the model. That’s where prompt engineering comes in. Prompt engineering entails the act of optimizing the input given to large language models (LLMs) to produce the most effective responses. 

Tips for paths to success

Here is a list of tips we recommend you use when inputting a prompt: 

1. Use formatting
2. Adopt a persona
3. Ask it to think step-by-step
4. Ask it to do one thing at a time
5. Don't ask it to combine disparate responses into one common response
6. Instruct the model to ask for clarification

These models are attempting to predict the next token, so it is important to give them context in the form of tokens to work with. If they are given a short prompt, that gives the models too few tokens upon which it is able to create new information from the context. Longer prompts are ideal, within the model's context window, as you can usually generate much better results. It's a balancing act between prompt length and prompt performance (maximizing per-token utility).

If you’re using a third party model, you can also implement some guardrails to restrict the paths that the model will take and the suggestions that it will provide. With some APIs and models, you can differentiate between user prompt vs. system prompt, and you can have system prompts that influence model behavior outside of the normal Human/Bot loop.

What if I get back something I didn’t want? 

You can run few shot or zero shot learning scenarios. This involves giving the model a series of example inputs and outputs that you want the model to emulate. From those tokens, it will learn how it’s supposed to predict the outcome of subsequent features. For instance, you could give the model a list of items and instruct it to return the list back in the specific format you are looking for. 

When prompting the models, ensure that you are using common formatting tools such as colons, quotes, and even markdown. A common scenario that we have encountered involves translating natural language to SQL or other query languages like, an  AWS OpenSearch query. We give it an example and put the query with the templated names in with three backticks (```) to help the LLM identify it as code and not text. It is important to ensure that things are formatted appropriately.

Another common technique is to ask the models to think step-by-step. LLMs need tokens about which to reason with, so it’s important to give it more tokens to generate, within a context window. You give the model specific instructions to think step-by-step about the prompts you are providing and ask it to seek approval before pursuing a task, and if it is not absolutely confident, it should prompt you for more information. 

Occasionally, conversations will also get stuck down the wrong path, so how do you correct it? Different models will behave differently. If a model has gone down a wrong path, you can ask it occasionally to correct itself and it will go and treat the full context of the situation of the prompt as if it has made a mistake along the way and it will attempt to predict what the next version is going to be. However, a lot of models won’t be able to do that as once they've gone down the wrong path, they're stuck in a local minima, so it's sometimes easier to start over with a new prompt.

Adopting a persona

Another key recommendation around prompt engineering involves asking the LLM to adopt a persona when given a prompt, however you have to be careful asking for persona’s that have limited or no context. When you tell the model to adopt a persona around which it has very little data in the model, it will give an answer worse than if you were to ask it to reply in a manner similar to an average human. This is because the model relies on the token generation within its existing model space and it has a data set that is extremely small for something like “world’s foremost expert on beanie babies”, and there have likely been reward modeling or other techniques utilized to deselect those paths as they often give incorrect results for the more common array of questions. To get better results, it is recommended to ask the models to behave as an expert or as an rather than as the best expert in the world.

Next Steps

Prompt engineering & Generative AI are still relatively new and there is still plenty of room to dig in and learn.

Are you exploring ways to take advantage of Analytical or Generative AI in your organization? Partnered with AWS, Caylent's data engineers have been implementing AI solutions extensively and are also helping businesses develop AI strategies that will generate real ROI. For some examples, take a look at our Generative AI offerings.