Insights by Grzegorz Luptowski, Backend Developer
Artificial Intelligence Search
In an era where users expect simplified interfaces and instant access to data, creating search experiences that feel effortless is anything but simple. Behind every search lies a complex software engineering challenge: delivering fast, relevant results while keeping the user experience intuitive.
Traditional search systems have long relied on carefully defined queries and deep domain knowledge. Simplifying that process to a single search field or a single click, while routing user input through language models, opens the door to user judgments and personal interpretations. This changes what users expect from the search process. Before any interaction with the search tool takes place, users often form expectations about what the system can do and what kinds of results it should return. These expectations are shaped by their previous experiences with similar search systems, and those experiences can have a significant influence on how they evaluate the results.
In this article, we’ll take a closer look at how the search process works in modern search systems, with a focus on vector search, and how the process itself is refined to improve result quality.
The user’s perspective
Let’s start by looking, in a simplified way, at how a user actually carries out a vector search.
The system’s perspective
We already know what things look like from the user’s side, but have you ever wondered what this looks like from a software engineering standpoint? Here we need to understand the concept of vector data representation, which is central to how search works.
In AI-based search, a vector (more precisely, an embedding, i.e. a vector representation) is a numerical representation of text, image, or audio data in the form of a list of floating-point numbers in a multidimensional space. An embedding is a vector of floating-point numbers, and the distance between two vectors measures their semantic resemblance: small distances indicate high similarity, and large distances indicate low similarity.
Once vectorized, data can be easily stored in vector databases, which enables fast, statistical comparison across various dimensions. The comparison process itself looks as follows:
A brief explanation
- Input data is passed into the embedding model.
- The model converts it into a vector.
- Vectors with similar meaning lie close to one another in that space.
- The user’s query is converted into a vector using the same method.
- The system computes cosine similarity between the query vector and the document vectors.
- Results are sorted from most to least similar.
Vector Search in Practice
Modern embeddings, produced by transformer-based models, encode a meaningful amount of semantic context. This is what gives vector search its advantage over purely lexical retrieval methods such as BM25 or TF-IDF. Rather than relying solely on exact keyword matches, vector search retrieves documents based on semantic similarity, making it particularly effective for searching large patent collections and surfacing relevant results that traditional keyword-based approaches may overlook.
It is important, however, to distinguish semantic similarity from true understanding or reasoning. An embedding model does not reason about a query, verify facts, or interpret intent in the way a human—or a large language model performing explicit reasoning—can. Instead, it creates a statistical, geometric representation of meaning and evaluates how closely documents relate to one another within that semantic space. This enables fast and highly effective retrieval of the most relevant candidates, providing a strong foundation for semantic patent search at scale.
Beyond Vector Search
While vector search delivers substantial improvements in retrieval quality, modern production search systems often combine multiple techniques to further refine results for specific use cases.
For example, hybrid search combines dense (vector) and sparse (keyword-based) retrieval to leverage the strengths of both approaches. Reranking models can then reorder the retrieved documents based on more advanced relevance evaluation, and in some scenarios the results can be further enhanced through Retrieval-Augmented Generation (RAG).
These techniques are rarely used in isolation. Instead, modern search systems combine them as complementary building blocks, selecting the most appropriate approach depending on the search task, data characteristics, and desired user experience.
1. RAG
Combines vector search with response generation by an LLM: the model generates an answer based on documents that were actually retrieved, rather than relying solely on its internal knowledge.
2. Fine-tuning
“Fine-tuning” actually covers two distinct processes that are often lumped together, and it’s worth telling them apart:
- Embedding model fine-tuning, the model that generates vectors is further trained on domain-specific pairs (e.g. query–relevant document pairs) so that it places semantically related domain content closer together in vector space. This directly improves retrieval quality, i.e. which documents get found in the first place.
- LLM fine-tuning, the language model that generates the final answer is further trained on domain-specific data or instructions so it better matches the terminology, tone, or task format expected in that domain. This improves generation quality, i.e. how the answer is phrased, but does not, by itself, change what gets retrieved.
In practice, teams often need only one of the two, and conflating them can lead to fine-tuning the wrong component for the problem at hand.
3. Reranking
An additional step after the initial search: a fast but less precise bi-encoder returns a set of candidates, and a slower cross-encoder re-sorts the results.
4. Hybrid Search
Combines vector search (dense retrieval, semantics) with lexical search such as BM25/TF-IDF (sparse retrieval, exact keyword matching). In production, pure vector search is rarely used on its own, dense retrieval handles synonyms and intent well, but struggles with rare terms, proper nouns, serial numbers, or domain-specific abbreviations, exactly the cases where sparse retrieval excels.
The most common way to combine both result lists is Reciprocal Rank Fusion (RRF), a simple, parameter-free algorithm that sums the inverse of a document’s rank position in each list, instead of comparing incompatible score scales (cosine similarity vs. BM25 score). This avoids the need to manually tune weights between the two methods.
Hybrid search can also be paired with reranking (technique #3): the dense+sparse fusion produces a broad, relevant set of candidates, and the cross-encoder performs the final, precise sort.
| Technique | What it does | Advantages | Limitations |
| RAG | Injects retrieved documents into the LLM’s context | Reduces hallucinations, easy knowledge updates, no training required | Quality depends on retrieval, adds latency |
| Fine-tuning | Further trains the embedding model (better retrieval) and/or the LLM (better generation) on proprietary data | Learns domain-specific terminology/jargon | Costly, risk of “forgetting,” hard to update, and easy to apply to the wrong component (retrieval vs. generation) |
| Reranking | Re-sorts results with a more precise model | Significant relevance boost at low cost | Can’t fix a poor first-stage retrieval |
| Hybrid Search | Combines dense (vector) and sparse (BM25/TF-IDF) retrieval, typically fused via RRF | Handles both semantic queries and exact-match terms that pure vector search misses | Infra complexity, fusion tuning still needed in edge cases |
Summary
Modern search systems are becoming increasingly sophisticated, with different approaches suited to different data environments, use cases, and user needs. While the user experience may appear simple—a single search field where users enter their query—multiple technologies and processes work behind the scenes to retrieve and rank relevant results. Choosing the right approach requires balancing factors such as speed, scalability, and result quality.
For patent search, this balance is especially important. Users need solutions that can handle large volumes of data while helping them discover relevant documents efficiently. Semantic search approaches add another layer to traditional search methods by identifying connections based on meaning rather than relying only on exact keyword matches.
If you’d like to see how this approach works in practice, explore the AI Search Suite in our platform. Start with a natural language prompt to discover relevant patents, or select an existing publication to find patents with similar technical features. The AI Search Suite helps you uncover meaningful connections across patent data without requiring you to anticipate every possible keyword variation. The current solution allows users to boost productivity by relying on relevant search results delivered in a short amount of time, letting them focus on their patents and work instead.
