Vector databases: chronicle of a foretold death - Generative AI France

🌈 Abstract

The article discusses the emergence of "Vector Databases" as the "memory" of Large-Language-Models (LLMs), and why they are more of a feature than a full-fledged product and already outdated. It explains the concept of vector representations in LLMs, the use of vector search in Retrieval Augmented Generation (RAG), and the limitations of vector databases compared to more comprehensive search engine solutions like Elasticsearch.

🙋 Q&A

[01] Vector Databases and Semantic Search

1. What are vector databases and how do they relate to LLMs?

• Vector databases store the vector representations (embeddings) of text, which capture the semantic meaning of the text. These vector representations are produced by the internal workings of LLMs.
• Vector search in these databases allows for semantic search, which is considered the best semantic search technique available so far.

2. What are the key use cases for vector search in LLMs?

• One of the main use cases is Retrieval Augmented Generation (RAG), which involves retrieving relevant documents from a database to provide specific answers to user questions, overcoming the high retraining cost of LLMs on new data.
• Vector search is used to find the most relevant documents in the database for a given question.

3. What are the limitations of vector search and vector databases?

• Vector search alone is not sufficient for real-world search applications, as documents also have important metadata (e.g., title, author, permissions) that should be considered in the retrieval process.
• Vector databases focus solely on the vector search functionality, while building a comprehensive search engine requires addressing additional challenges like query planning, scalability, and data integrity.

[02] Comparison to Elasticsearch

1. How does Elasticsearch compare to vector databases?

• Elasticsearch, as an established search engine, supports vector type fields and proximity queries for semantic search, in addition to a wide range of other query types and scoring functions.
• Elasticsearch also has proven scalability and performance, potentially outperforming specialized vector databases in some benchmarks.

2. What advantages does Elasticsearch have over vector databases?

• Elasticsearch supports a richer set of search features beyond just vector search, such as keyword search, scoring functions, and Boolean query operators.
• Elasticsearch has a more mature ecosystem with better SDKs, documentation, and hosting options, compared to the newer vector database startups.
• Elasticsearch has a proven track record of handling large-scale search requirements, while the sustainability of vector database startups is more uncertain.

3. What is the author's overall perspective on vector databases vs. Elasticsearch?

• The author suggests that vector databases are too narrowly focused on a single feature (vector search), while the real challenges of building a comprehensive search engine lie elsewhere, and are better addressed by more mature solutions like Elasticsearch.