Skip to main content

​
Overview

HyDE, or Hypothetical Document Embeddings, is an approach that enhances retrieval accuracy by generating hypothetical answers based on user queries. In the RAG framework, HyDE is used to improve search relevance in cases where direct query results might lack sufficient context. By leveraging hypothetical answers, generated by an LLM, HyDE can enrich query embeddings, ensuring that retrieved documents are aligned more closely with the user’s intent.

​
How HyDE works in the RAG framework

1

Generate Hypothetical Answer

When a user query is processed, a hypothetical answer is generated by the LLM, reflecting the most likely information that the user seeks. This answer serves as an enriched form of the query.For example: A user asks, “What are the quarterly revenue trends?” A hypothetical answer, such as “The quarterly revenue has increased over time,” is generated to add context.
2

Embed and Retrieve

The hypothetical answer is embedded using vectorization, creating a more contextually informed query vector. This vector is then used in a vector similarity search, enhancing the relevance of retrieved documents.
3

Refinement with Re-ranking

After the initial retrieval, the results are re-ranked using the cross-encoder model to prioritize documents that match the query’s intent, further refining the list of results.

​
The Agent role

  1. Agent-Generated Hypothetical Answer: When an AI agent processes a user query, it may generate a hypothetical answer that represents a likely response. This hypothetical answer is provided as the hypothetical_answer argument in the knowledge base retrieval tool, enriching the search query with additional context.
  2. Embedding and Retrieval: The hypothetical answer, along with the user query, is embedded as a vector. This enriched query vector is then used to perform a similarity search in the vector database (Weaviate). By incorporating the hypothetical answer, the system can retrieve documents that are semantically closer to the ideal response, even if they don’t directly match the keywords in the user’s initial query.

​
Sample Setup in Knowledge Base Retrieval Tool

export const retrieveKnowledge = {
    type: "function",
    function: {
        name: "retrieve_knowledge",
        description: "Always use this tool to access the knowledge base and respond based on provided information. Always use all the information provided from this tool.",
        parameters: {
            type: 'object',
            properties: {
                user_input: {
                    type: 'string',
                    description: 'The user input to retrieve knowledge from the knowledge base',
                },
                hypothetical_answer: {
                    type: 'string',
                    description: 'Given the user query, provide a detailed and informative answer. Use the same language as the user input.',
                },
                intention: {
                    type: "string",
                    description: "The short English description of the intention of the user.",
                }
            },
            required: ['user_input', 'hypothetical_answer'],
        },
    },
};