The popularity of vector search has surged in recent years, primarily driven by the rise of AI. The technology behind vector search is quite complex, but implementing vector search is not that difficult. In principle, any experienced developer can get started with vector search.
Before you start using vector search, the most important question is: When do you need vector search, and when is it better than your current traditional search technology? At the moment, it does not appear that vector search will completely replace traditional search technology, for various reasons, which we will discuss later. We spoke at length with Steve Kearns, General Manager for Search Solutions at Elastic, to determine for whom vector search is suitable.
Also read: Elastic: The power of search enables AI
What is vector search?
Vector search differs fundamentally from traditional search methods. Whereas traditional search engines search for literal words and terms, vector search is about understanding meaning. “Vector search is an approach that searches for the most relevant or related content by using the actual meaning of words and phrases,” Kearns explains.
The process works very differently from traditional search methods. Vector search uses specialized language models (not the large LLMs such as ChatGPT, but targeted embedding models) to convert text into numerical representations, known as vectors, which capture the meaning of the text. This enables search engines to make connections between different terminologies. If you search for “car,” the system can also find documents that mention “vehicle” or “motor vehicle,” even if those exact terms do not appear.
The power of multimodality
A major advantage of vector search is the ability to search different media types. “If you want to search a database of images, how do you make it indexable for a search engine? It doesn’t contain any words,” says Kearns. Vector embeddings make it possible to convert images directly into numerical representations (vectors), allowing you to search for visual content using text.
This principle is already being applied by companies such as Vimeo. They use both transcripts with timestamps at the sentence level and screenshots at keyframes from videos, and make them searchable with embeddings. Adobe has been using Elasticsearch for image search for years, long before vector search became known through AI.
When should you consider vector search?
Many organizations already have search engines, whether on their websites or internally, for accessing employee information. When should you consider adding vector search?
If semantic meaning is crucial, vector search can be a good solution. This is the case when users search for the same information using different words, or when a better search query can lead to increased revenue. A large e-commerce platform could potentially achieve 1 or 2 percent more revenue by applying vector search. The application of vector search is therefore immediately measurable. Suppose someone searches for a blue dress, a yellow cabinet, or a tile with a specific pattern. In that case, vector search makes it much easier to display similar products, especially if all product images are in the vector search database.
Or what if you need to search documents from different sources, not just text, but also images and videos? Vector search can also offer a solution for domain-specific knowledge.
However, vector search is not a direct replacement or competitor for traditional search. If you are searching very specifically for a particular document in which specific terms are used, a traditional search will give you that result immediately, while vector search will apply semantic search terms and possibly return many more results that you will then have to sift through to find what you were ultimately looking for.
Hybrid search works best: traditional and vector search together
Ultimately, Kearns argues that Elastic is convinced that a hybrid form of search is best. The best search engine is capable of combining traditional search methods and vector search. This provides the best performance. “In almost all cases, you get the best performance by combining traditional and vector search.”
It ensures that you can capture results that match exactly and present them first, while at the same time establishing semantic connections and presenting those results as well. This almost always provides an accurate overall picture, without requiring the user to search through the results again.
Implementation challenges
Vector search does add extra complexity. Documents or texts must be divided into chunks, then run through embedding models, and finally indexed efficiently. Elastic uses HNSW (Hierarchical Navigable Small World) indexing for this.
To keep things from getting too complex, Elastic has chosen to integrate it into its existing search solution. It is an additional data type that can be stored in a column alongside existing data. This also makes hybrid search much easier. However, this is not so simple with every vector search provider.
Cost optimization
Vector search is considerably more expensive than traditional search methods. The problem lies in the memory-intensive nature of vector operations. Elastic has developed a solution for this called BBQ (Better Binary Quantization).
The different quantization techniques in Elasticsearch compress 32-bit floats to 8-bit integers or even 1-bit representations, significantly reducing memory usage. “The great thing is that we can query that compressed data directly without decompression,” explains Kearns. This not only saves costs but also often improves performance through hardware optimizations. However, you do need a special instance with a language model to generate the vectors, which are a lot more expensive than normal instances.
Vector search performance in practice
Modern vector search is basically no slower than traditional search methods. “In most cases, vector search is as fast as traditional lexical search, and sometimes even faster,” says Kearns. Low-level optimizations down to CPU instructions (SIMD, AVX) allow vector operations to run at the same speed as text searches.
GPU’s are required for embedding generation, converting documents and text into vectors, but searches can run efficiently on CPUs.
Integration within the AI ecosystem
Vector databases have become virtually indispensable for modern AI applications. Elastic offers integrations with popular frameworks such as LangChain and Llama Index. Elastic also recently launched an MCP (Model Context Protocol) server for seamless integration with AI agents. We understand from Kearns that there are a few more integrations on the roadmap.
Pitfalls and realistic expectations
Organizations that want to switch to vector search immediately should not expect miracles. The result ultimately depends on the available data. As with AI, the same rule applies, garbage in means garbage out. “Language models are not magic. Vector search is not magic either,” Kearns emphasizes. Success depends on several factors:
- The right embedding model for your domain
- Clean, well-structured data
- Appropriate security controls
- Filtering and refinement capabilities
Many organizations that initially started with pure vector databases have transitioned to platforms that offer greater flexibility. “We see many organizations that started with vector-only search and now say that it is not good enough.” At Elastic, they are therefore moving much more towards hybrid search. This is also something we observe at Google, where the Google search engine serves as a hybrid search environment that combines multiple forms of search in the background.
The future is multimodal and specialized
Innovation in embedding models continues, particularly in the development of multimodal applications that combine text, images, audio, and video within a single search experience. General-purpose models are getting better and better, making domain-specific fine-tuning less necessary.
Conclusion
Many organizations can benefit from vector search, provided the data is of high quality and there is a good use case. The message for IT professionals and developers is therefore simple: start with a use case.
Are you dealing with semantic search queries, multimodal content, or AI-driven applications? Then vector search is very valuable. However, implement it as part of a hybrid strategy, rather than as a replacement for existing search capabilities.
Start small, experiment with hybrid search, and scale based on results. Vector databases are powerful tools, but as Kearns concludes: “You need a search engine that contains vectors, not just a vector database.”
The future of enterprise search lies not in choosing between traditional or vector search, but in intelligently combining both to deliver the best possible results.