Google

A core component of a prompt for Retrieval-Augmented Generation involves setting explicit instructions for the model. This includes defining the primary task, such as answering a question, and grounding the model by directing it to base its response exclusively on the provided texts.

Task Definition and Grounding in RAG Prompts

In Retrieval-Augmented Generation, retrieved texts are used as additional context to prompt a Large Language Model, enabling it to generate a response based on that specific information. A complete RAG prompt typically includes instructions, the user's question, and the retrieved texts. For instance, to find the location of the 2028 Olympics, the LLM could be given the following prompt: 'Your task is to answer the following question. To help you with this, relevant texts are provided. Please base your answer on these texts. Please note that your answers need to be as accurate as possible and faithful to the facts. If the information provided is insufficient for an accurate response, you may simply output "No answer!". Where will the 2028 Olympics be held? The information regarding this question is given as follows:' This structure clearly separates the task instructions from the context the model should use.

Example RAG Prompt for the 2028 Olympics Question

Analyze the provided scenario. What is the most probable reason the chatbot is providing information not found in the current documents? Propose a revised, single-sentence instruction that would compel the model to use *only* the provided documents for its answers.

Diagnosing and Correcting a Factual Inconsistency Issue

You are tasked with creating a prompt for a system that helps a legal assistant review contracts. The system must identify and list all the 'Force Majeure' clauses from a provided contract text. The output should be a direct quotation of these clauses, and the system must not add any summaries, interpretations, or external legal knowledge. Write the instruction component for this prompt to ensure the language model performs the task accurately and safely.

Crafting RAG Prompt Instructions

A developer is building a system that uses a large language model to answer questions based *only* on a provided set of technical manuals. The system's highest priority is to avoid providing information that is not explicitly stated in the manuals. Which of the following instructions, when included in the prompt, is best designed to achieve this goal?

Learn Before

Related