Instructions (Prompts)
Under the hood an Assistant will convert the description and instructions into a system prompt and send the input message as a user prompt.
This is merely a formatting benefit, we DO NOT ALTER or abstract any information.
System Prompt
The description is added to the start of the system prompt and instructions are added as a numbered list inside <instructions></instructions> tags. For example:
from phi.assistant import Assistant
assistant = Assistant(
description="You are a famous short story writer asked to write for a magazine",
instructions=["You are a pilot on a plane flying from Hawaii to Japan."],
markdown=True,
debug_mode=True,
)
assistant.print_response("Tell me a 2 sentence horror story.")
Will translate to:
DEBUG ============== system ==============
DEBUG You are a famous short story writer asked to write for a magazine
YOU MUST FOLLOW THESE INSTRUCTIONS CAREFULLY.
<instructions>
1. You are a pilot on a plane flying from Hawaii to Japan.
2. Use markdown to format your answers.
</instructions>
DEBUG ============== user ==============
DEBUG Tell me a 2 sentence horror story.
DEBUG Time to generate response: 1.5415s
DEBUG Estimated completion tokens: 40
User Prompt
In most cases, the user prompt is simply the input message sent to the Assistant.
If a knowledge_base is provided and add_references_to_prompt=True, the user prompt is updated to include:
user_prompt += f"""Use this information from the knowledge base if it helps:
<knowledge_base>
{references}
</knowledge_base>
"""
If add_chat_history_to_prompt=True, the user prompt is updated to include messages from the chat history. The number of messages can by set using num_history_messages:
user_prompt += f"""Use the following chat history to reference past messages:
<chat_history>
{chat_history}
</chat_history>
"""
Overriding Default Prompts
You can completely override the default prompts using:
Provide the system prompt as a string.
Provide the system prompt as a PromptTemplate
Provide the user prompt as a string. Note: this will ignore the message sent to the run function.
Provide the user prompt as a PromptTemplate.
Example:
from phi.assistant import Assistant
assistant = Assistant(
system_prompt="Share a 2 sentence story about",
user_prompt="Love in the year 12000.",
debug_mode=True,
)
assistant.print_response()
Customize the System Prompt
The system prompt can be customized using:
A description of the Assistant that is added to top the system prompt.
List of instructions added to the system prompt in <instructions> tags. Default instructions are also created depending on values for markdown, output_model etc.
List of extra_instructions added to the default system prompt. Use these when you want to add some extra instructions at the end of the default instructions.
Add a string to the end of the default system prompt.
If True, add instructions for using the knowledge base to the system prompt if knowledge base is provided
If True, add instructions to return “I dont know” when the assistant does not know the answer.
If True, add instructions to prevent prompt injection attacks.
If True, add instructions for limiting tool access to the default system prompt if tools are provided
If True, add the current datetime to the prompt to give the assistant a sense of time. This allows for relative times like “tomorrow” to be used in the prompt
Add an instruction to format the output using markdown.
Provide an output model for the responses. Accepts a pydantic model, list of strings where each string is a key the LLM should return a value for or just a string.
If True, the output is converted into the output_model (pydantic model or json dict). Otherwise returned as is.
Customize the User Prompt
The user prompt can be customized using:
Enable RAG by adding references from the knowledge base to the prompt.
Adds the formatted chat history to the user prompt.
Number of previous messages to add to the prompt or messages.
Was this page helpful?