Google

A prompt can be designed to provide a Large Language Model with multiple avenues for formulating an answer. Such a hybrid prompt might combine instructions for using external tools with pre-retrieved information. For instance, the prompt could begin with instructions like, 'Your task is to answer the following question. You may use external tools, such as web search, to assist you,' followed by the user's query, 'Where will the 2028 Olympics be held?'. It can then introduce the provided context by stating, 'The information regarding this question is given as follows:'. This structure gives the model the flexibility to synthesize an answer from the given texts or to perform its own search if the context is inadequate.

Example of a Hybrid Prompt Combining RAG and Tool Use

An AI assistant is tasked with answering a user's question. Its prompt is structured to include both a set of pre-retrieved documents related to the query and instructions that permit it to perform a live web search. What is the primary strategic advantage of this combined approach?

Given the following scenario, propose a modification to the AI assistant's operational instructions that would allow it to provide more accurate, real-time answers.

Improving an AI Travel Assistant's Accuracy

You are designing a prompt for an AI assistant to answer the user query: 'What is the current stock price of Company XYZ and how has it performed over the last year?' The system has access to pre-retrieved annual performance reports for Company XYZ. It can also use an external tool to get the live, current stock price. Write a single, effective prompt that instructs the assistant on how to use both the provided information and the external tool to formulate a comprehensive answer.

Learn Before

Related