RAG Techniques

Decoupling in Retrieval-Augmented Generation (RAG) systems means separating the query embedding process from retrieving search results, allowing each part to be optimized independently. Enhancements can be made to the query embedding algorithms and retrieval mechanisms separately, leading to potential improvements in overall system performance.

For example, a company is improving its customer support chatbot. User queries, such as “How can I reset my password?”, are embedded using a model that was fine-tuned using company-specific interactions, which optimizes the query embedding process. Independently, the document retrieval is optimized to prioritize documents based on relevance, recency, and user feedback. Optimizing each process independently allows for more accurate and contextually rich responses, significantly improving the customer support chatbot’s effectiveness.

The retrieval process in Retrieval-Augmented Generation (RAG) systems can be improved by including additional context, such as incorporating preceding and following chunks along with the retrieved chunk from a search. By including additional text, the model gains a broader context around the most relevant chunk, leading to more informed and accurate responses. Additional context helps capture important information that may not be present in the isolated chunk, resulting in higher-quality responses.

For example, a customer support chatbot responds to queries such as “How do I reset my password?”. If the most relevant chunk contains “Step 1: Go to login page” but doesn’t contain information in surrounding chunks, such as “Step 2: Click ‘Forgot Password’…” and “Password issues can be resolved by…”, the resultant answer may not provide the most comprehensive support information. With this additional context, a wider range of password issues can be resolved in the initial response.

Contextual Query Retrieval (CQR) is an advanced technique that enhances the retrieval process in Retrieval-Augmented Generation (RAG) systems by incorporating the entire conversation history when fetching relevant information. Using conversation history, CQR ensures the retrieved information is more relevant and coherent, leading to contextually appropriate responses. This technique significantly enhances the quality of interactions in applications like chatbots and virtual assistants, where conversation context evolves.

Hypothetical Document Embeddings (HyDE) is an advanced technique in Retrieval-Augmented Generation (RAG) systems that uses a language model to generate a hypothetical answer based on the query. This hypothetical answer is then embedded and used for the similarity search. Unlike standard RAG workflows that embed the user’s query directly, HyDE introduces an intermediary step of generating a hypothetical answer before embedding. Using a hypothetical answer can lead to more accurate retrievals by embedding terms that align closely with those found in relevant documents.

For example, if a user inputs a query, “How can I improve my Wi-Fi signal at home?”, the standard RAG workflow will directly embed this query. The embedded query is used to search a database of support documents, which retrieves the most relevant document. For HyDE, a language model generates a hypothetical answer based on the query, such as “You can improve your Wi-Fi signal by placing the router in a central location, minimizing interference from other devices, using a Wi-Fi extender, and ensuring your router’s firmware is up-to-date.” HyDE embeds the hypothetical answer for the similarity search, which can lead to a more comprehensive and accurate response by capturing additional relevant details through the hypothetical answer.

Fusion Search is an advanced Retrieval-Augmented Generation (RAG) technique designed to manage complex queries by breaking them down into smaller, more manageable subqueries. Each subquery is processed separately through the RAG system to retrieve precise and relevant information. The results from each subquery are then combined to form a thorough and cohesive response. This method mimics human problem-solving strategies, making complex problems easier to tackle.

For example, if the original query is “Explain the causes and effects of climate change on polar ecosystems,” an LLM model generates relevant subqueries such as:

• “What are the primary causes of climate change?”
• “How does climate change affect polar ecosystems?”
• “What are the effects of rising temperatures on polar wildlife?”
• “How does melting ice impact polar environments?”

Each subquery is then processed using the RAG system. For instance, the answer to the first subquery might be: “The primary causes of climate change include the burning of fossil fuels, deforestation, and industrial activities that increase greenhouse gas emissions.” All the subquery answers are then combined and fed into an LLM with the original query in the prompt to produce a comprehensive answer.

Prompts should be straightforward and to the point, avoiding unnecessary complexity and ambiguity. Use directives that tell the model what to do (“do this”) rather than what not to do (“don’t do this”). Positive language helps guide the model more effectively. Clear and concise prompts reduce the likelihood of misunderstanding or misinterpretation by the model. Positive language ensures that the model focuses on desired behaviors and outcomes, leading to better-quality responses.