AIR Search: Semantic search with AI – Increased efficiency through context-based information retrieval

It is becoming increasingly difficult for companies to identify relevant content efficiently. Traditional full-text searches have structural and conceptual limitations: they deliver results based on exact keywords without taking into account the meaning of the content or the semantic context. With the increasing digitization of business processes—especially in knowledge management, IT documentation, and contract management—there is a growing need for intelligent search solutions that not only compare character strings but also understand content.

AIR Search addresses precisely this problem with a modern approach based on artificial intelligence. The semantic search engine uses technologies such as natural language processing (NLP), embeddings, and vector-based similarity calculation to interpret user queries in terms of content and deliver contextually relevant answers. Instead of relying on rigid keyword searches, AIR Search translates words and sentences into so-called “embeddings” – numerical formats that represent the meaning and context of terms. This enables the search engine to recognize that ‘notebook’ and “laptop” mean the same thing in many cases. With the help of vector search, AIR Search can then search particularly efficiently for information with similar content – even if it is formulated completely differently from the original query. This not only makes the search more accurate, but also much more intuitive for the user. This article highlights the technological basics, practical benefits, and integration options of this AI-powered search solution.

1. From keyword search to semantic information retrieval

Traditional search engines – such as those found in file directories, intranets, or document management systems – are usually based on exact word matches. The term entered is compared with indexed text content, and hits are often sorted by frequency or position of the term in the document. This approach works well for unique terms, but fails with synonyms, unclear wording, or context-dependent meanings.

In contrast, semantic search takes a meaning-oriented approach: The search query is first analyzed linguistically and converted into a semantic vector space using modern language models. Instead of searching for character strings, the system filters for concepts with similar content. This makes it possible to deliver relevant results even when the terms in the documents differ, such as in the case of synonyms or related technical terms.

What makes AIR Search so accurate? - Embeddings & vector search explained simply

The technology behind AIR Search combines efficient data structuring with modern language models – it sounds complex, but is surprisingly intuitive to use.

Embeddings – meaning instead of word form

Words, sentences, or entire documents are translated by AI into so-called vectors – numbers that represent not the word itself, but its meaning. This allows the system to recognize that “notebook” and “laptop” often mean the same thing.

Vector search – intelligent finding

Instead of searching for exact terms, AIR Search compares the meaning of your query with the contents of the database. This also works with synonyms or differently worded questions, delivering more relevant results.

Your advantage as a user or decision-maker

• Less frustration due to irrelevant hits
• Comprehensible answers – even for complex questions
• Faster orientation in large amounts of data

2. Technological basis: How AIR Search understands content

The core technology of AIR Search is based on modern AI approaches, particularly in the field of NLP. It uses transformer models such as BERT (Bidirectional Encoder Representations from Transformers) or SBERT (Sentence-BERT), which are capable of analyzing the semantic structure of a text and transferring it to a high-dimensional vector space.

2.1 Embeddings and vector search

Every sentence, phrase, and document is mapped as a vector—i.e., a mathematical point in space. The proximity of these vectors to each other reflects their semantic similarity. When a search query is entered, it is also translated into a vector and compared with the stored document vectors using cosine similarity or other distance metrics. Results with a low distance, i.e., high semantic similarity, are displayed first.

Diese Technik ermöglicht eine fehlertolerante, kontextbasierte Suche, die über rein formale Übereinstimmungen hinausgeht. AIR Search nutzt hierbei optimierte Vektordatenbanken wie FAISS oder OpenSearch mit ANN-Indexierung, um auch bei großen Datenmengen performante Suchzeiten sicherzustellen. ANN steht für „Approximate Nearest Neighbor“ – eine Methode, bei der nicht jede mögliche Übereinstimmung geprüft wird, sondern besonders wahrscheinliche Treffer blitzschnell eingegrenzt werden. So findet AIR Search in Sekundenbruchteilen relevante Ergebnisse, auch wenn Milliarden von Vektoren durchsucht werden müssen. Der Vorteil: hohe Suchgeschwindigkeit bei nur minimalem Genauigkeitsverlust – ein perfekter Kompromiss zwischen Qualität und Effizienz.

2.2 Wie bewertet AIR Search Ergebnisse und wie entstehen gespeicherte Vektoren?

Die Vektoren, die AIR Search verwendet, werden bei der Indexierung der zugrunde liegenden Inhalte erzeugt. Das können Dokumente, Webseiten, Datenbankeinträge oder Support-Tickets sein. Mithilfe unseres vortrainierten KI-Modells (z. B. BERT oder Sentence Transformers) werden diese Inhalte in die genannten semantischen Vektoren umgewandelt. Die Suchintelligenz kann aber individuell auf spezielle Unternehmensanforderungen durch spezifische Trainingsdaten für unsere vortrainierten Modelle zugeschnitten werden.

Effizient und ganz ohne Training wird aber bei Änderungen im Datenbestand, wie z.B. neue Dokumente, Webseiten usws., der Index automatisch aktualisiert.

Ein wichtiger Punkt für den operativen Betrieb: Ein einmal vektorisierter Inhalt bleibt stabil. Der erzeugte Vektor verändert sich nicht mehr – auch wenn später durch viele Änderungen weitere Inhalte in den Index aufgenommen werden. Das sorgt für Konsistenz, Nachvollziehbarkeit und hohe Effizienz im laufenden Systembetrieb.

2.3 Kontextualisierung und Sprachverständnis

Darüber hinaus integriert AIR Search spezialisierte Sprachmodelle, die in der Lage sind, Domänenwissen zu berücksichtigen. Ein juristischer Text wird so anders verarbeitet als ein technischer Handbuchauszug. Durch kontextsensitives Fine-Tuning kann AIR Search auf die spezifische Fachsprache einzelner Branchen abgestimmt werden – ein entscheidender Vorteil gegenüber generischen KI Suchmaschinen.

2.4 Retrieval-Augmented Generation (RAG) gibt konkrete Antworten

Zusätzlich zur semantischen Suche, welche eine Erweiterung bzw. Verbesserung der Schlüsselwort-basierten Suche darstellt, bietet AIR Search auch die Anbindungsmöglichkeit von RAG und damit zusätzliche Verarbeitungsmöglichkeiten von Datenbanken, Dokumenten, oder Knowledge bases.

Während die Semantische Suche ausformulierte Nutzerfragen durch das Ranking von Suchergebnissen beantwortet, wird im RAG-Prozess ein generatives LLM (Large Language Model) mit einer semantischen Suche kombiniert, um die Antworten des LLMs mit Informationen aus Kundendokumenten anzureichern. LLMs sind große Sprachmodelle wie Llama 3.2 oder ChatGPT 4.0, welche mit riesigen Mengen an Daten trainiert wurden. RAG kann dabei genauere und fundiertere Antworten liefern als ein generisches LLM, da zusätzlich zu den Trainingsdaten des Sprachmodells Dokumente, aus welchen mittels semantischer Suche Informationen eingeholt werden, zur Generierung der Antwort hinzugezogen werden.

Der RAG-Prozess gestaltet sich folgendermaßen: Erst wird die Nutzereingabe als Suchanfrage verwendet, mit welcher die semantische Suche auf Nutzerdokumente durchgeführt wird. Die Top-Suchergebnisse, also jene Dokumente welche bei der semantischen Suche als Ergebnisse aufscheinen, werden nun als Kontext für die Antwort des Sprachmodells verwendet. Schließlich gibt das RAG-System eine Fließtextantwort aus, welche die Inhalte der Top-Suchergebnisse miteinbezieht. Die zur Antwortgenerierung verwendeten Dokumente können außerdem als Quellen angeführt werden, was zu einer vertrauenswürdigeren Antwort mit höherer Nachvollziehbarkeit führt.

This technology enables error-tolerant, context-based searches that go beyond purely formal matches. AIR Search uses optimized vector databases such as FAISS or OpenSearch with ANN indexing to ensure high-performance search times even with large amounts of data. ANN stands for “Approximate Nearest Neighbor” – a method in which not every possible match is checked, but particularly likely hits are narrowed down at lightning speed. This enables AIR Search to find relevant results in a fraction of a second, even when billions of vectors have to be searched. The advantage: high search speed with only minimal loss of accuracy – a perfect compromise between quality and efficiency.

2.2 How does AIR Search evaluate results and how are stored vectors created?

The vectors used by AIR Search are generated when the underlying content is indexed. This can be documents, websites, database entries, or support tickets. Using our pre-trained AI model (e.g., BERT or Sentence Transformers), this content is converted into the aforementioned semantic vectors. However, the search intelligence can be customized to specific business requirements by providing specific training data for our pre-trained models.

Efficiently and without any training, the index is automatically updated when changes are made to the data, such as new documents, web pages, etc.

An important point for operational use: Once content has been vectorized, it remains stable. The generated vector does not change, even if further content is added to the index later due to numerous changes. This ensures consistency, traceability, and high efficiency in ongoing system operation.

2.3 Contextualization and language understanding

In addition, AIR Search integrates specialized language models that are capable of taking domain knowledge into account. A legal text is processed differently than an excerpt from a technical manual. Through context-sensitive fine-tuning, AIR Search can be tailored to the specific technical language of individual industries – a decisive advantage over generic AI search engines.

2.4 Retrieval-Augmented Generation (RAG) provides concrete answers

In addition to semantic search, which is an extension or improvement of keyword-based search, AIR Search also offers the option of connecting RAG, thereby providing additional processing options for databases, documents, or knowledge bases.

While semantic search answers formulated user questions by ranking search results, the RAG process combines a generative LLM (large language model) with semantic search to enrich the LLM's answers with information from customer documents. LLMs are large language models such as Llama 3.2 or ChatGPT 4.0, which have been trained with huge amounts of data. RAG can provide more accurate and informed answers than a generic LLM because, in addition to the language model's training data, documents from which information is retrieved using semantic search are also used to generate the answer.

The RAG process is as follows: First, the user input is used as a search query to perform a semantic search on user documents. The top search results, i.e., the documents that appear as results in the semantic search, are now used as context for the language model's response. Finally, the RAG system outputs a continuous text response that incorporates the content of the top search results. The documents used to generate the response can also be cited as sources, resulting in a more trustworthy response with greater traceability.