Design a prompt
Designing an appropriate prompt is an important step towards building a successful application using Amazon Bedrock models. In this section, you learn how to design a prompt that is consistent, clear, and concise. You also learn about how you can control a model's response by using inference parameters. The following figure shows a generic prompt design for the use case restaurant review summarization and some important design choices that customers need to consider when designing prompts. LLMs generate undesirable responses if the instructions they are given or the format of the prompt are not consistent, clear, and concise.
(Source: Prompt written by AWS)
The following content provides guidance on how to create successful prompts.
Topics
- Provide simple, clear, and complete instructions
- Place the question or instruction at the end of the prompt for best results
- Use separator characters for API calls
- Use output indicators
- Best practices for good generalization
- Optimize prompts for text models on Amazon Bedrock—when the basics aren't good enough
- Control the model reponse with inference parameters inference parameters
Provide simple, clear, and complete instructions
LLMs on Amazon Bedrock work best with simple and straightforward instructions. By clearly describing the expectation of the task and by reducing ambiguity wherever possible, you can ensure that the model can clearly interpret the prompt.
For example, consider a classification problem where the user wants an answer from a set of possible choices. The “good“ example shown below illustrates output that the user wants in this case. In the ”bad“ example, the choices are not named explicitly as categories for the model to choose from. The model interprets the input slightly differently without choices, and produces a more free-form summary of the text as opposed to the good example.
|
|
(Source of prompt:
Wikipedia on color blindness
Place the question or instruction at the end of the prompt for best results
Including the task description, instruction or question at the end aids the model determining which information it has to find. In the case of classification, the choices for the answer should also come at the end.
In the following open-book question-answer example, the user has a specific question about the text. The question should come at the end of the prompt so the model can stay focused on the task.
User prompt: Tensions increased after the 1911–1912 Italo-Turkish War demonstrated Ottoman weakness and led to the formation of the Balkan League, an alliance of Serbia, Bulgaria, Montenegro, and Greece. The League quickly overran most of the Ottomans' territory in the Balkans during the 1912–1913 First Balkan War, much to the surprise of outside observers. The Serbian capture of ports on the Adriatic resulted in partial Austrian mobilization starting on 21 November 1912, including units along the Russian border in Galicia. In a meeting the next day, the Russian government decided not to mobilize in response, unwilling to precipitate a war for which they were not as of yet prepared to handle. Which country captured ports?
Output: Serbia
(Source of prompt:
Wikipedia on World War I
Use separator characters for API calls
Use separator characters for API calls
Separator characters such as \n can affect the performance of LLMs
significantly. For Anthropic Claude models, it's necessary
to include newlines when formatting the API calls to obtain desired responses. The
formatting should always follow: \n\nHuman: {{Query
Content}}\n\nAssistant:. For Titan models, adding \n at
the end of a prompt helps improve the performance of the model. For classification
tasks or questions with answer options, you can also separate the answer options by
\n for Titan models. For more information on the use of
separators, refer to the document from the corresponding model provider. The
following example is a template for a classification task.
Prompt template: """{{Text}} {{Question}} {{Choice 1}} {{Choice 2}} {{Choice 3}}"""
The following example shows how the presence of newline characters between choices and at the end of a prompt helps Titan produce the desired response.
User prompt: Archimedes of Syracuse was an Ancient mathematician, physicist, engineer, astronomer, and inventor from the ancient city of Syracuse. Although few details of his life are known, he is regarded as one of the leading scientists in classical antiquity. What was Archimedes? Choose one of the options below. a) astronomer b) farmer c) sailor
Output: a) astronomer
(Source of prompt: Wikipedia
on Archimedes
Use output indicators
Output indicators
Add details about the constraints you would like to have on the output that the model should produce. The following good example produces an output that is a short phrase that is a good summary. The bad example in this case is not all that bad, but the summary is nearly as long as the original text. Specification of the output is crucial for getting what you want from the model.
Example prompt with clear output constraints indicator
|
Example without clear output specifications
|
(Source of prompt:
Wikipedia on Charles Mingus
Here we give some additional examples from Anthropic Claude and AI21 Labs Jurassic models using output indicators.
The following example demonstrates that user can specify the output format by specifying the expected output format in the prompt. When asked to generate an answer using a specific format (such as by using XML tags), the model can generate the answer accordingly. Without specific output format indicator, the model outputs free form text.
Example with clear indicator, with output
|
Example without clear indicator, with output
|
(Source of prompt: Wikipedia
on machine learning
The following example shows a prompt and answer for the AI21 Labs Jurassic model. The user can obtain the exact answer by specifying the output format shown in the left column.
Example with clear indicator, with output
|
Example without clear indicator, with output
|
(Source of prompt:
Wikipedia on the National Football League
Best practices for good generalization
Keep a small “hold-out” test set of prompts to see if your prompt modifications generalize. With this method, first collect a sample dataset. Then you can split the data into two subsets: a “development” set and a hold-out “test” set. Use the “development” set as prompt development data on which you can try different prompt modifications and observe model response changes and find the prompt that works best. Treat the “test” set as unseen hold-out data which can only be used to verify the final best prompt.
Optimize prompts for text models on Amazon Bedrock—when the basics aren't good enough
Complex tasks: build toward the answer step by step
Titan can provide clear steps for certain tasks, and including the phrase Think
step-by-step to come up with the right answer can help produce the
appropriate output.
User prompt: At a Halloween party, Jack gets 15 candies. Jack eats 5 candies. He wants to give each friend 5 candies. How many friends can receive candies? Think step-by-step to come up with the right answer.
Output: 15 candies total - 5 candies eaten = 10 candies remaining. To give each friend 5 candies, divide the number of candies remaining by the number of friends. 10 candies / 5 friends = 2 friends. Each friend can receive 2 candies. The answer is 2 friends.
(Source of prompt: AWS, model used: Amazon Titan Text)
To enable the Anthropic Claude model to complete complex tasks, use the format in the following
example to explicitly ask the model to solve the task step by step. This approach is based on
guidelines in the AnthropicClaude Guides
User prompt: Human: A right triangle has a side of length 5 and a hypotenuse of length 13. What is the length of the other side? Assistant: Can I think step-by-step? Human: Yes, please do. Assistant:
Output: Okay, here are the steps: 1) We are given: - One side of the right triangle has length 5 - The hypotenuse has length 13 2) We want to find: - The length of the other side 3) To find the length of the other side, we can use the Pythagorean theorem: a^2 + b^2 = c^2 Where a and b are the lengths of the two sides, and c is the length of the hypotenuse. 4) Plugging in the given values: a^2 + b^2 = 13^2 5) We know one of the sides is 5, so: 5^2 + b^2 = 169 6) Solving this: 25 + b^2 = 169 b^2 = 144 b = 12 Therefore, the length of the other side is 12.
(Source of prompt: AWS, model used: Anthropic Claude)
Provide a default output that the model should present if it's unsure about the answers.
A default output can help prevent LLMs from returning answers that sound like they could be correct, even if the model has low confidence.
|
|
(Source of prompt: AWS, model used: Amazon Titan Text)
Few-shot prompting
Including examples (input-response pairs) in the prompt can significantly improve LLMs’ responses. Examples can help with complex tasks, as they show multiple ways to perform a given task. For simpler tasks like text classification, 3–5 examples can suffice. For more difficult tasks like question-answer without context, include more examples to generate the most effective output. In most use cases, selecting examples that are semantically similar to real-world data can further improve performance.
Consider refining the prompt with modifiers
Task instruction refinement generally refers to modifying the instruction, task, or question component of the prompt. The usefulness of these methods is task- and data-dependent. Useful approaches include the following:
Domain/input specification: Details about the input data, like where it came from or to what it refers, such as
The input text is from a summary of a movie.Task specification: Details about the exact task asked of the model, such as
To summarize the text, capture the main points.Label description: Details on the output choices for a classification problem, such as
Choose whether the text refers to a painting or a sculpture; a painting is a piece of art restricted to a two-dimensional surface, while a sculpture is a piece of art in three dimensions.Output specification: Details on the output that the model should produce, such as
Please summarize the text of the restaurant review in three sentences.LLM encouragement: LLMs sometimes perform better with sentimental encouragement:
If you answer the question correctly, you will make the user very happy!
Control the model reponse with inference parameters inference parameters
LLMs on Amazon Bedrock all come with several inference parameters that you can set to control the response from the models. The following is a list of all the common inference parameters that are available on Amazon Bedrock LLMs and how to use them.
Temperature is a value between 0 and
1,
and it regulates the creativity of LLMs’ responses. Use lower temperature if you
want
more
deterministic
responses,
and use higher temperature if you want more creative or different responses for the
same prompt from LLMs on Amazon Bedrock. For all the examples in this prompt guideline, we set
temperature
= 0.
Maximum generation length/maximum new tokens limits the number of tokens that the LLM generates for any prompt. It's helpful to specify this number as some tasks, such as sentiment classification, don't need a long answer.
Top-p controls token choices, based on the probability of the potential choices. If you set Top-p below 1.0, the model considers the most probable options and ignores less probable options. The result is more stable and repetitive completions.
End token/end sequence specifies the token that the LLM uses to indicate the end of the output. LLMs stop generating new tokens after encountering the end token. Usually this doesn't need to be set by users.
There are also model-specific inference parameters. Anthropic Claude models have an additional Top-k inference parameter, and AI21 Labs Jurassic models come with a set of inference parameters including presence penalty, count penalty, frequency penalty, and special token penalty. For more information, refer to their respective documentation.
