Retrieval-augmented generation (RAG) is a technique that enables generative artificial intelligence (Gen AI) models to retrieve and incorporate new information.^[1] It modifies interactions with a large language model (LLM) so that the model responds to user queries with reference to a specified set of documents, using this information to supplement information from its pre-existing training data.^[2] This allows LLMs to use domain-specific and/or updated information.^[2][3] Use cases include providing chatbot access to internal company data or generating responses based on authoritative sources.^[4]

RAG improves large language models (LLMs) by incorporating information retrieval before generating responses. Unlike traditional LLMs that rely on static training data, RAG pulls relevant text from databases, uploaded documents, or web sources.^[1] According to Ars Technica, "RAG is a way of improving LLM performance, in essence by blending the LLM process with a web search or other document look-up process to help LLMs stick to the facts." This method helps reduce AI hallucinations, which have led to real-world issues like chatbots inventing policies or lawyers citing nonexistent legal cases.^[5]

By dynamically retrieving information, RAG enables AI to provide more accurate responses without frequent retraining. According to IBM, "RAG also reduces the need for users to continuously train the model on new data and update its parameters as circumstances evolve. In this way, RAG can lower the computational and financial costs of running LLM-powered chatbots in an enterprise setting."^[1]

Beyond efficiency gains, RAG also allows LLMs to include source references in their responses, enabling users to verify information by reviewing cited documents or original sources. This can provide greater transparency, as users can cross-check retrieved content to ensure accuracy and relevance.^[4]

In June 2024, Ars Technica reported, "But LLMs aren’t humans, of course. Their training data can age quickly, particularly in more time-sensitive queries. In addition, the LLM often can’t distinguish specific sources of its knowledge, as all its training data is blended together into a kind of soup." In 2023, during its launch demonstration, Google’s Bard provided incorrect information about the James Webb Space Telescope, an error that contributed to a $100 billion decline in Alphabet’s stock value.^[5]

RAG systems do not inherently verify the credibility of the sources they retrieve from, which can lead to misleading or inaccurate responses. AI systems can generate misinformation even when pulling from factually correct sources if they misinterpret the context. MIT Technology Review gives the example of an AI-generated response stating, "The United States has had one Muslim president, Barack Hussein Obama." The model retrieved this from an academic book titled Barack Hussein Obama: America’s First Muslim President? and misinterpreted the content, generating a false claim.^[2]

Retrieval-Augmented Generation (RAG) is a method that allows large language models (LLMs) to retrieve and incorporate additional information before generating responses. Unlike LLMs that rely solely on pre-existing training data, RAG integrates newly available data at query time. Ars Technica states, "The beauty of RAG is that when new information becomes available, rather than having to retrain the model, all that’s needed is to augment the model’s external knowledge base with the updated information."^[5]

The BBC describes "prompt stuffing" as a technique within RAG, in which relevant context is inserted into a prompt to guide the model’s response. This approach provides the LLM with key information early in the prompt, encouraging it to prioritize the supplied data over pre-existing training knowledge.^[6]

Retrieval-Augmented Generation (RAG) enhances large language models (LLMs) by incorporating an information-retrieval mechanism that allows models to access and utilize additional data beyond their original training set. AWS states, "RAG allows LLMs to retrieve relevant information from external data sources to generate more accurate and contextually relevant responses" (indexing).^[4] This approach reduces reliance on static datasets, which can quickly become outdated. When a user submits a query, RAG uses a document retriever to search for relevant content from available sources before incorporating the retrieved information into the model’s response (retrieval). Ars Technica notes that "when new information becomes available, rather than having to retrain the model, all that’s needed is to augment the model’s external knowledge base with the updated information" (augmentation).^[5] By dynamically integrating relevant data, RAG enables LLMs to generate more informed and contextually grounded responses (generation). IBM states that "in the generative phase, the LLM draws from the augmented prompt and its internal representation of its training data to synthesize an engaging answer tailored to the user in that instant.^[1]

Typically, the data to be referenced is converted into LLM embeddings, numerical representations in the form of a large vector space. RAG can be used on unstructured (usually text), semi-structured, or structured data (for example knowledge graphs).^[3] These embeddings are then stored in a vector database to allow for document retrieval.

Given a user query, a document retriever is first called to select the most relevant documents that will be used to augment the query.^[2] This comparison can be done using a variety of methods, which depend in part on the type of indexing used.^[1][3]

The model feeds this relevant retrieved information into the LLM via prompt engineering of the user's original query.^[4] Newer implementations (as of 2023) can also incorporate specific augmentation modules with abilities such as expanding queries into multiple domains and using memory and self-improvement to learn from previous retrievals.^[3]

Finally, the LLM can generate output based on both the query and the retrieved documents.^[2][7] Some models incorporate extra steps to improve output, such as the re-ranking of retrieved information, context selection, and fine-tuning.^[3]

Improvements to the basic process above can be applied at different stages in the RAG flow.

These methods center around the encoding of text as either dense or sparse vectors. Sparse vectors, used to encode the identity of a word, are typically dictionary length and contain almost all zeros. Dense vectors, used to encode meaning, are much smaller and contain far fewer zeros. Several enhancements can be made to the way similarities are calculated in the vector stores (databases).

• Performance can be improved with faster dot products, approximate nearest neighbors, or centroid searches.^{[clarification needed]}
• Accuracy can be improved with Late Interactions.^{[clarification needed][8]}
• Hybrid vectors: dense vector representations can be combined with sparse one-hot vectors in order to use the faster sparse dot products rather than the slower dense ones.^[9] Other^{[clarification needed]} methods can combine sparse methods (BM25, SPLADE) with dense ones like DRAGON.

These methods focus on improving the quality of hits from the vector database:

• pre-train the retriever using the Inverse Cloze Task.^[10]
• progressive data augmentation. The method of Dragon samples difficult negatives to train a dense vector retriever.^[11]
• Under supervision, train the retriever for a given generator. Given a prompt and the desired answer, retrieve the top-k vectors, and feed those vectors into the generator to achieve a perplexity score for the correct answer. Then minimize the KL-divergence between the observed retrieved vectors probability and LM likelihoods to adjust the retriever.^[12]
• use reranking to train the retriever.^[13]

Retro language model for RAG. Each Retro block consists of Attention, Chunked Cross Attention, and Feed Forward layers. Black-lettered boxes show data being changed, and blue lettering shows the algorithm performing the changes.

By redesigning the language model with the retriever in mind, a 25-time smaller network can get comparable perplexity as its much larger counterparts.^[14] Because it is trained from scratch, this method (Retro) incurs the high cost of training runs that the original RAG scheme avoided. The hypothesis is that by giving domain knowledge during training, Retro needs less focus on the domain and can devote its smaller weight resources only to language semantics. The redesigned language model is shown here.

It has been reported that Retro is not reproducible, so modifications were made to make it so. The more reproducible version is called Retro++ and includes in-context RAG.^[15]

Chunking involves various strategies for breaking up the data into vectors so the retriever can find details in it.

Different data styles have patterns that correct chunking can take advantage of.

Three types of chunking strategies are:

• Fixed length with overlap. This is fast and easy. Overlapping consecutive chunks helps to maintain semantic context across chunks.
• Syntax-based chunks can break the document up into sentences. Libraries such as spaCy or NLTK can also help.
• File format-based chunking. Certain file types have natural chunks built in, and it's best to respect them. For example, code files are best chunked and vectorized as whole functions or classes. HTML files should leave <table> or base64 encoded <img> elements intact. Similar considerations should be taken for pdf files. Libraries such as Unstructured or Langchain can assist with this method.

Rather than using documents as a source to vectorize and retrieve from, Knowledge Graphs can be used. One can start with a set of documents, books, or other bodies of text, and convert them to a knowledge graph using one of many methods, including language models. Once the knowledge graph is created, subgraphs can be vectorized, stored in a vector database, and used for retrieval as in plain RAG. The advantage here is that graphs has more recognizable structure than strings of text and this structure can help retrieve more relevant facts for generation. Sometimes this approach is called GraphRAG.^{[citation needed]}

Sometimes vector database searches can miss key facts needed to answer a user's question. One way to mitigate this is to do a traditional text search, add those results to the text chunks linked to the retrieved vectors from the vector search, and feed the combined hybrid text into the language model for generation. ^{[citation needed]}

If the external data source is large, retrieval can be slow.^[16]

RAG is not a complete solution to the problem of hallucinations in LLMs. According to Ars Technica, "It is not a direct solution because the LLM can still hallucinate around the source material in its response."^[5]

While RAG improves the accuracy of large language models (LLMs), it does not eliminate all challenges. One limitation is that while RAG reduces the need for frequent model retraining, it does not remove it entirely. Additionally, LLMs may struggle to recognize when they lack sufficient information to provide a reliable response. Without specific training, models may generate answers even when they should indicate uncertainty. According to IBM, this issue can arise when the model lacks the ability to assess its own knowledge limitations.^[1]

RAG systems may retrieve factually correct but misleading sources, leading to errors in interpretation. In some cases, an LLM may extract statements from a source without considering its context, resulting in an incorrect conclusion. Additionally, when faced with conflicting information, RAG models may struggle to determine which source is accurate and may combine details from multiple sources, producing responses that merge outdated and updated information in a misleading way. According to the MIT Technology Review, these issues occur because RAG systems may misinterpret the data they retrieve.^[2]