Qdrant’s BM42 Algorithm Boosts RAG Efficiency, Challenging Tech Giants in Enterprise AI Race

Vector database company Qdrant has developed a new search algorithm called BM42 to make retrieval augmented generation (RAG) more efficient and cost-effective, as more companies look to incorporate RAG into their technology stack.

Qdrant’s BM42 algorithm aims to improve RAG efficiency: BM42 is designed to provide vectors to companies working on new search methods, focusing on hybrid search that combines semantic and keyword search:

• BM42 is an update to the BM25 algorithm used by traditional search platforms to rank document relevance in search queries, which assumes documents have enough size to calculate statistics.
• With RAG often using vector databases that store data as mathematical metrics for easy data matching, BM42 uses a language model to extract information from documents as tokens, which are then scored or weighted to rank their relevance to the search question.
• This approach allows Qdrant to pinpoint the exact information needed to answer a query, making it more suitable for working with smaller chunks of information in RAG compared to BM25.

Alternative methods for hybrid search and RAG: While BM42 aims to improve upon BM25, it is not the first method developed for this purpose:

• Splade (Sparse Lexical and Expansion model) is another option that works with a pre-trained language model to identify relationships between words and include related terms that may not be identical between the search query text and the referenced documents.
• However, Qdrant’s co-founder and CTO, Andrey Vasnetsov, argues that BM42 is a more cost-efficient solution compared to Splade, which can be expensive and slow due to the large models and computational requirements.

Growing interest in RAG for enterprise AI: RAG is becoming a hot topic in enterprise AI as companies seek ways to use generative AI models and map them to their own data:

• RAG could provide more accurate and real-time information from company data to employees and other users.
• Major tech companies like Microsoft and Amazon now offer infrastructure for cloud computing clients to build RAG applications, while OpenAI recently acquired Rockset to strengthen its RAG capabilities.
• However, it is important to note that RAG, being a language model, can still be prone to hallucinations, as highlighted in a recent Stanford study on AI legal research tools.

Analyzing the implications and challenges: As companies increasingly adopt RAG and seek more efficient and cost-effective solutions, innovations like Qdrant’s BM42 algorithm could play a significant role in shaping the future of enterprise AI and search capabilities. However, the industry must also address the challenges associated with RAG, such as the potential for hallucinations and the need for robust data management and integration. As the competitive landscape evolves, with major players like Microsoft, Amazon, and OpenAI investing heavily in RAG, it will be crucial for companies to carefully evaluate their options and consider the trade-offs between performance, cost, and reliability when implementing these technologies.