Improving Retrieval Augmented Generation (RAG) Systems with Qdrant’s BM42 Algorithm
As more companies seek to enhance their technology stack with retrieval augmented generation (RAG) systems, the demand for more efficient and cost-effective methods is on the rise. Qdrant, a vector database company, has developed a new search algorithm called BM42, which aims to address these needs.
Qdrant, founded in 2021, created BM42 to offer vectors to companies working on innovative search methods. The company’s goal is to provide customers with a hybrid search approach that combines semantic and keyword search capabilities. By leveraging vector databases, which store data as mathematical metrics for easy data matching, Qdrant aims to enhance RAG systems.
Andrey Vasnetsov, co-founder and chief technology officer of Qdrant, explained in an interview with VentureBeat that BM42 is an update to the traditional search algorithm BM25. While BM25 assumes documents have enough size to calculate statistics, RAG systems work with smaller chunks of information. In this context, BM42 uses a language model to extract information from documents and convert it into tokens. These tokens are then scored or weighted by the algorithm to determine their relevance to the search query. This allows Qdrant to precisely identify the necessary information to answer a query.
While BM42 offers an improved approach to hybrid research and RAG, it is not the only alternative to BM25. Splade, which stands for Sparse Lexical and Expansion model, is another option in this space. Splade utilizes a pre-trained language model that can identify relationships between words and include related terms that may not be exact matches between the search query and the referenced documents.
Although other vector database companies use Splade, Vasnetsov believes that BM42 provides a more cost-efficient solution. Splade can be expensive due to the large size of the models and the computational resources required. BM42 offers a more affordable and faster alternative.
RAG systems are gaining popularity in enterprise AI as companies seek ways to leverage generative AI models and align them with their own data. RAG has the potential to deliver more accurate and real-time information from company data to employees and other users. Major players like Microsoft and Amazon are already offering infrastructure for building RAG applications in their cloud computing services. In fact, OpenAI recently acquired Rockset to strengthen its RAG capabilities.
However, it’s important to note that RAG systems, despite their advantages, are still language models that can be prone to hallucinations. While they allow users to ground AI models in company data, there is a risk of generating inaccurate or misleading information.
In conclusion, Qdrant’s BM42 algorithm offers a promising solution to improve RAG systems by providing more efficient and cost-effective search capabilities. As the demand for hybrid research and RAG continues to grow, companies can explore alternative methods like Splade but should consider the advantages of BM42, particularly its affordability and speed. With RAG systems becoming increasingly essential in enterprise AI, companies have the opportunity to leverage generative AI models to unlock valuable insights from their own data. However, it is crucial to be cautious of potential pitfalls and ensure the accuracy and reliability of the generated information.