Prompt engineering techniques

This guide doesn't go in-depth into the mechanics behind the message structure for Chat Completions. If you aren't familiar with interacting with Chat Completions models programmatically, we recommend reading our how-to guide on the Chat Completion API first.

System message

The system message is included at the beginning of the prompt and is used to prime the model with context, instructions, or other information relevant to your use case. You can use the system message to describe the assistant’s personality, define what the model should and shouldn’t answer, and define the format of model responses.

The example below, shows a sample system message and the resulting model response:

Some other examples of system messages are:

• “Assistant is a large language model trained by OpenAI.”
• “Assistant is an intelligent chatbot designed to help users answer technical questions about Azure OpenAI Service. Only answer questions using the context below and if you're not sure of an answer, you can say "I don't know".
• “Assistant is an intelligent chatbot designed to help users answer their tax related questions.”
• “You're an assistant designed to extract entities from text. Users will paste in a string of text and you'll respond with entities you've extracted from the text as a JSON object. Here's an example of your output format:

{  
   "name": "",
   "company": "",
   "phone_number": ""
}

An important detail to understand is that even if you instruct a model in the system message to answer I don't know when unsure of an answer this doesn't guarantee that the request will be honored. A well designed system message can increase the likelihood of a certain outcome, but it's still possible that an incorrect response could be generated that contradicts the intent of the instruction in the system message.

Few-shot learning

A common way to adapt language models to new tasks is to use few-shot learning. In few-shot learning, a set of training examples is provided as part of the prompt to give additional context to the model.

When using the Chat Completions API, a series of messages between the User and Assistant (written in the new prompt format), can serve as examples for few-shot learning. These examples can be used to prime the model to respond in a certain way, emulate particular behaviors, and seed answers to common questions.

The table above describes the basic prompt structure, but for a detailed explanation of the exact prompt format you'll need to refer to the Chat Completion guide.

Non chat scenarios

While the Chat Completion API is optimized to work with multi-turn conversations, it also can be used for non chat scenarios. For example, for a sentiment analysis scenario, you might use the following prompt:

Start with clear instructions

The sequence information appears in the prompt matters. This is because GPT style models are built in a certain way that defines how they process the input. Our research suggests that telling the model the task you want it to do at the beginning of the prompt, before sharing additional contextual information or examples, can help produce higher-quality outputs.

Repeat instructions at the end

Models can be susceptible to recency bias, which in this context means that information at the end of the prompt might have more significant influence over the output than information at the beginning of the prompt. Therefore, it's worth experimenting with repeating the instructions at the end of the prompt and evaluating the impact on the generated response.

Prime the output

This refers to including a few words or phrases at the end of the prompt to obtain a model response that follows the desired form. For example, using a cue such as “Here’s a bulleted list of key points:\n- ” can help make sure the output is formatted as a list of bullet points.

In the above prompt, the text One possible search query is: primes the model to produce a single output. Without this cue the model produces several search queries as output.

Add clear syntax

Using clear syntax for your prompt—including punctuation, headings, and section markers—helps communicate intent and often makes outputs easier to parse.

In the example below, separators (--- in this case) have been added between different sources of information or steps. This allows the use of --- as a stopping condition for generation. In addition, section headings or special variables are presented in uppercase to differentiate them.

If you're not sure what syntax to use, consider using Markdown or XML. The models have been trained on a large quantity web content in XML and Markdown, which might provide better results.

Break the task down

Large language models (LLMs) often perform better if the task is broken down into smaller steps. For example, in the search query prompt referenced earlier, the prompt can be restructured so that the model is first instructed to extract relevant facts, and then instructed to generate search queries that can be used to verify those facts.

Notice the use of clear syntax to differentiate the sections and prime the output. In this simple example, breaking the task down from one to two steps isn't very dramatic, but when trying to do this for a larger piece of text with many factual claims, breaking the task down can make a significant difference.

Use of affordances

Sometimes we can get the model to use an affordance instead of relying on its own parameters for information and answers. Search, for example, can be an affordance to help mitigate against fabricated answers, and to get up-to-date information.

One simple way to use an affordance is to stop generation once the affordance calls are generated by the model, then paste the outcomes back into the prompt. Below is an example of what the follow-up call could be after doing the SEARCH calls above. Notice how we pasted the search results into the prompt and replaced the previous SEARCH calls.

Chain of thought prompting

This is a variation on the break the task down technique. Instead of splitting a task into smaller steps, in this approach, the model response is instructed to proceed step-by-step and present all the steps involved. Doing so reduces the possibility of inaccuracy of outcomes and makes assessing the model response easier.

Specifying the output structure

Using your prompt to specify the structure of the output can have a significant effect on the nature and quality of the results. Sometimes system message inputs as “only write true facts” or “don't fabricate information” might not be adequate mitigations. Instead, asking for the model response to also include citations can help result in a reduction of the prevalence of incorrect responses.

If you instruct the model to cite the source material when it makes statements, those statements are much more likely to be grounded. Asking for citations makes it so that the model must make two errors every time it generates a response: the first error is the fabricated response, and the second is the bad citation. Note that the closer the citation is to the text it supports, the shorter the distance ahead the model needs to anticipate the citation, which suggests that inline citations are better mitigations for false content generation than citations at the end of the content.

Similarly, if you ask the model to extract factual statements from a paragraph, it might extract compound statements such as ‘X is doing Y AND Z’ (which can be harder to verify). This can be avoided by specifying an output structure such as (entity1, relationship, entity2).

The examples below show the use of citations and directing the model response to fit a defined structure.

Temperature and Top_p parameters

Changing the temperature parameter changes the output of the model. The temperature parameter can be set between 0 and 2. A higher value, for example 0.7 makes the output more random and produce more divergent responses, while a lower value, like. 0.2, makes the output more focused and concrete. A fictional story could be generated using a higher temperature. Whereas to generate a legal document it's recommended to use a much lower temperature. Top_probability is another parameter and is similar to Temperature in that it also controls the randomness of the model response, but it does so in a different way. The general recommendation is to alter one of these two parameters at a time, not both.

Provide grounding context

One of the most effective ways to provide reliable answers is to give the model data to draw its responses from (grounding data). If your use case relies on up-to-date, reliable information and isn't purely a creative scenario, we strongly recommend providing grounding data. In general, the closer you can get your source material to the final form of the answer you want, the less work the model needs to do, which means there's less opportunity for error. Here's an example where the system is provided a recent blog describing the launch of GPT-4 in Azure OpenAI Service, and asked to name some of its early customers.

¹ Text from Azure OpenAI GPT-4 launch blog.

Next steps

• Learn more about Azure OpenAI.
• Learn more about the Chat Completions API.
• Azure OpenAI Samples.

Start with clear instructions

The sequence information is added into the prompt matters. Our research suggests that telling the model the task you want it to do at the beginning of the prompt, before sharing additional contextual information or examples, can help produce higher-quality outputs.

In the first example below, a key part of the instructions is shared only at the end, and the resulting completion is incorrect.

In this next example, we give the exact statement we want to check (“Several sources mention a chance of another eruption”) before we give the snippet context. This allows the token representations extracted from the snippets to be tailored to the statement we're checking. Now the resulting model response is accurate:

Repeat the instructions at the end

Models can be susceptible to recency bias, which means that information at the end of the prompt might have more significant influence over the output than information at the beginning of the prompt. Therefore, it's worth experimenting with repeating the instructions at the end of the prompt and evaluating the impact for your use case.

We saw in the earlier example how repeating the instruction at the beginning and the end of the prompt leads to a higher likelihood of getting an accurate model response.

Prime the output

This refers to including a few words or phrases at the end of the prompt to obtain a model response that follows the desired form. For example, using a cue such as “Here’s a bulleted list of key points:\n- ” can help make sure the output is formatted as a list of bullet points.

In the above prompt, the text "One possible search query is:" primes the model to produce an output in the form of a search query. This technique can help remove hedges that the model might typically add, such as “One possibility is...”.

Add clear syntax

Using clear syntax for your prompt—including punctuation, headings, and section markers—helps communicate intent and often makes outputs easier to parse.

In the example below, separators (--- in this case) have been added between different sources of information or steps. This allows the use of --- as a stopping condition for generation. In addition, section headings or special variables are presented in uppercase to differentiate them.

If you’re not sure what syntax to use, consider using markdown or XML, since large language models have been trained on a lot of web content in XML or markdown.

Break the task down

Models often perform better if the task is broken down into smaller steps. For example, in the search query prompt referenced earlier, the prompt can be restructured so that the model is first instructed to extract relevant facts, and then instructed to generate search queries that can be used to verify those facts.

Notice the use of clear syntax to differentiate the sections and prime the output. In this simple example, breaking the task down from one to two steps is not very dramatic, but when trying to do this for a larger piece of text with many factual claims, breaking the task down can make a significant difference.

Use of affordances

Sometimes we can get the model to use an affordance instead of relying on its own parameters for information and answers. Search, for example, can be an affordance to help mitigate against fabricated answers, and to get up-to-date information.

One simple way to use an affordance is to stop generation once the affordance calls are generated by the model, then paste the outcomes back into the prompt. Below is an example of what the follow-up call could be after doing the SEARCH calls above. Notice how we pasted the search results into the prompt and replaced the previous SEARCH calls.

Chain of thought prompting

This is a variation on the break the task down technique. Instead of splitting a task into smaller steps, in this approach, the model response is instructed to proceed step-by-step and present all the steps involved. Doing so reduces the possibility of inaccuracy of outcomes and makes assessing the model response easier.

Meta prompts

If you find that the model response is not as desired, it can often be helpful to add a meta prompt that directly corrects the behavior. This is a directive prepended to the instruction portion of the prompt. Consider this example, where the computer seems to be overly aggressive:

While meta prompts can often help minimize unintentional bad outputs, keep in mind they aren't robust against adversarial attacks since they are vulnerable to prompt injection.

Further, conditional statements explaining what the model should do if it encounters a particular circumstance are often more effective than just telling the model not to do something.

Few-shot learning

A very common way to adapt language models to new tasks is to use few-shot learning. In few-shot learning a set of training examples is provided in the prompt and then the model is asked to complete one or more unfinished examples. In the following example we use an instruction combined with few-shot learning to make up puns:

Choose your few-shot examples carefully and ensure they cover a variety of circumstances relevant to your scenario, including edge cases.

Also, as shared earlier, the large language models can exhibit a form of recency bias. This means that the order in which "few-shot" examples are provided to the model, matters. If you don't want this to matter, consider sampling multiple completions from prompts based on randomized orderings of the examples or list items.

Use quotes to generate a single sentence

If the desired model response is only a single simple sentence or string, we can end the completion guidance with a quote symbol ", and set " as a stop sequence.

Specifying output structure

Using your prompt to specify the structure of the output can have a significant impact on the nature and quality of the results. Sometimes meta prompts such as “only write true facts” or “do not fabricate information” might not be adequate mitigations. Instead, asking for the model response to also include citations can help result in a reduction of the prevalence of incorrect responses.

For example, if you force the LLM to cite the source material when it makes statements, those statements are much more likely to be grounded. Asking for citations makes it so that the model must make two errors every time it fabricates a response: the first error is the fabricated response, and the second is the bad citation. Note that the closer the citation is to the text it supports, the shorter the distance ahead the LLM needs to anticipate the citation, which suggests that inline citations are better mitigations for false content generation than citations at the end of the content.

Similarly, if you ask the model to extract factual statements from a paragraph, it might extract compound statements such as ‘X is doing Y AND Z’ (which might be harder to verify), which can be avoided by specifying an output structure such as (entity1, relationship, entity2).

Adjusting Temperature and Top_P parameters

Changing the temperature parameter allows models to have differing risk levels. A higher value, for example 0.7 will make the output more random and produce more divergent responses. While a lower value, like 0.2 will make the output more focused and concrete. For example, a fictional story could be generated using a higher temperature. Whereas to generate a legal document it's recommended to use a much lower temperature.

In the next example we test two model responses at different temperature settings of 0.1 and 0.9 respectively. While the evaluation of creative content such as the pieces of text below is a subjective exercise and opinions might vary, the difference in model response to the same prompt solely from changing the temperature parameter is apparent.

Top_probability is another parameter and is similar to temperature in that it also controls the randomness of the model response, but it does so in a different way. Experiment with different model responses by varying temperature and Top_P, but not both at once.

Next steps

• Learn more about Azure OpenAI.
• Get started with the ChatGPT model with the ChatGPT quickstart.
• For more examples, check out the Azure OpenAI Samples GitHub repository