Terms from A–Z
Agentic AI
Agentic AI refers to AI systems that autonomously pursue goals, create plans and execute actions – without requiring human input for every step. Unlike reactive chatbots, Agentic AI systems work with multiple specialized sub-agents that operate in parallel and in a coordinated manner.
Not to be confused with:
Traditional chatbots or rule-based bots that are limited to predefined responses. Agentic AI plans and acts; a chatbot only reacts.
Practical application:
In e-commerce, Agentic AI handles product consulting: analyzing queries, searching the catalog, asking follow-up questions and explaining recommendations – without manual rule maintenance.
Agentic Search
Agentic Search coordinates multiple specialised AI agents that translate purchase intent, requirements and constraints into a complete catalogue match across structured product data points per SKU. The result is not a top-k subset, but a complete, reasoned selection of all matching products.
Not to be confused with:
Semantic Search or RAG. Agentic Search uses no vector similarity, but attribute-based intent matching with a follow-up question mechanism.
Practical application:
In the online shop: The zero-result rate is significantly reduced because agents find matching product attributes even for unknown or ambiguous queries.
RAG (Retrieval-Augmented Generation)
RAG combines information retrieval with text generation. A language model receives relevant documents from a vector database and generates a response based on them. Goal: reduce hallucinations and incorporate current information.
Not to be confused with:
Agentic Search. RAG generates answers from documents; Agentic Search delivers product recommendations from structured catalog data without a vector index.
Practical application:
Insurance companies use RAG for FAQ answering from policy documents. The hallucination risk remains elevated, however, as generative outputs are not verified.
LLM (Large Language Model)
A Large Language Model is a neural network trained on large amounts of text that understands and generates language. LLMs form the foundation for systems like ChatGPT, Claude or Gemini. They generate probable continuations – they are not databases and do not verify facts.
Not to be confused with:
A search engine. LLMs generate text, they do not index URLs. They can hallucinate because they do not verify facts.
Practical application:
In e-commerce, LLMs serve as the language layer: they interpret customer queries and formulate responses. The actual product logic lies with specialized agents.
Embeddings
Embeddings are numerical vectors that represent the semantic content of texts, images or products. Similar content receives similar vectors – this enables similarity search. Generated by language models or specialized embedding models.
Not to be confused with:
Keywords. Keywords are exact terms; embeddings represent meaning. Two sentences with different words can have very similar embeddings.
Practical application:
In e-commerce: product descriptions are stored as embeddings to find semantically similar products during search – regardless of exact phrasing.
Vector Search
Vector Search finds documents or products based on similarity between embeddings, not based on exact keywords. Each query is converted to a vector; the index returns the most similar vectors.
Not to be confused with:
Agentic Search. Vector Search measures similarity; Agentic Search understands context and asks follow-up questions. Vector Search can fail with very specific attribute queries.
Practical application:
E-commerce shops use Vector Search for semantic product search. New products without sufficient description are poorly found – Agentic Search bypasses this through attribute matching.
Conversational Commerce
Conversational Commerce refers to trade via dialogue-based interfaces: chat, voice or messenger. Customers have a conversation, ask questions and receive product suggestions in dialogue – automated and scalable.
Not to be confused with:
Live chat. Live chat connects with humans; Conversational Commerce is automated and scales to millions of simultaneous conversations.
Practical application:
Shops with Agentic AI-based Conversational Commerce report shorter customer journeys and higher cart values through precise needs clarification.
AI Product Discovery
AI Product Discovery describes the process by which AI systems help customers discover relevant products – even when they do not know what they are looking for. Typical features: dialogue-based needs clarification, personalized suggestions, explanations.
Not to be confused with:
Traditional product search. Traditional search requires a known search term; AI Product Discovery actively guides the customer to suitable products.
Practical application:
Insurance providers use AI Product Discovery for tariff recommendations: the AI asks about needs, explains differences and recommends the appropriate tariff.
Hallucinations (AI)
Hallucinations refer to factually incorrect or fabricated statements that an AI language model produces with apparent certainty. They arise because LLMs maximize probabilities, not verify facts. Hallucinations are a systemic characteristic of generative models – not an error in the classical sense.
Not to be confused with:
Misinformation due to data problems. Hallucinations arise in the generation process itself, regardless of training quality.
Practical application:
In e-commerce, hallucinated product properties are a returns and liability risk. Agentic Search minimizes the risk by deriving recommendations directly from verified product attributes.
Explainable AI (XAI)
Explainable AI refers to AI systems whose decisions can be transparently explained. A recommendation is considered explainable if the system can justify which attributes or features led to the output.
Not to be confused with:
Transparent training. Explainability refers to the output, not the training process.
Practical application:
Ein Explain-Panel zeigt, warum ein Produkt empfohlen wurde: „Passend wegen: Gewicht < 1,5 kg, 15-Zoll-Display, langer Akkulaufzeit." Das stärkt Vertrauen und erfüllt BaFin-Anforderungen für Versicherungsprodukte.
Re-Ranking
Re-Ranking is a two-stage search process: First stage – fast pre-selection of candidates. Second stage – a model evaluates each candidate more precisely and re-sorts the list. Result: higher relevance than single-stage search.
Not to be confused with:
Sorting by filters. Filters reduce the search space rule-based; Re-Ranking evaluates semantic relevance between query and candidates.
Practical application:
E-commerce search systems use Re-Ranking to sort top-10 results by purchase probability. With Agentic Search, the agents perform this evaluation directly.
Query Intent
Query Intent bezeichnet die eigentliche Absicht hinter einer Suchanfrage. „Laptop für Homeoffice" signalisiert: leise, leicht, langer Akku – obwohl diese Attribute nicht explizit genannt werden. KI-Systeme, die Query Intent verstehen, liefern bessere Ergebnisse als rein lexikalische Suchen.
Not to be confused with:
Keywords. Keywords are words in the query; Query Intent is the underlying purchase intention. The same intent can be expressed through many different formulations.
Practical application:
Agentic Search detects intent, asks clarifying questions on ambiguity, and filters the catalogue by inferred attributes – without the customer needing to know product specifications.
Zero-Result (No Results)
A Zero-Result occurs when a search returns no results – even though matching products would be in the catalog. Causes: too narrow keyword matches, missing synonyms, insufficient intent understanding. Zero-Results cause purchase abandonment.
Not to be confused with:
An empty catalogue. Zero-Results arise from an inadequate search solution, not from missing products.
Practical application:
Traditional keyword searches produce Zero-Results with typos, synonyms or combined attributes. Agentic Search eliminates Zero-Results structurally: agents always find matching attributes.
POC (Proof of Concept)
A Proof of Concept is a structured test operation that shows whether a technology works under real conditions. In the AI context: integration into the real system environment, testing with real product data, measurement of specific KPIs – before the full rollout.
Not to be confused with:
Demo or simulation. Demos show features in sample systems; a POC runs on your own data and delivers measurable results as a basis for decision-making.
Practical application:
AGINITY AI delivers POCs in 48 hours: upload product catalog, configure agents, test live. Typical KPIs: zero-result rate, conversion, return rate.
GDPR-compliant AI
GDPR-compliant AI refers to AI systems that meet the requirements of the General Data Protection Regulation: data minimization, purpose limitation, transparency and processing within the EU legal framework. For AI, additionally: documentation obligation for automated decisions (Art. 22 GDPR).
Not to be confused with:
ISO 27001 certification or security audits. GDPR compliance concerns the handling of personal data, not primarily IT security.
Practical application:
Shops and insurance providers in the DACH region must ensure that AI recommendations are traceable and contestable. AGINITY AI processes all requests pseudonymously in German data centers.
AI Product Advisory
AI product consulting is dialogue-based consulting in which a system captures needs, selects suitable products and justifies the recommendation. The goal is not just a results list, but decision support. Particularly effective for products requiring explanation with many variants.
Not to be confused with:
Simple keyword search without context. Keyword searches deliver results lists; AI product consulting clarifies needs and explains the selection.
Practical application:
Hilft Shops und Versicherern, komplexe Produkte verständlich zu empfehlen und Kaufabbrüche zu reduzieren. Beispiel: „Ich brauche einen leisen Laptop für Homeoffice" → Rückfragen + begründete Auswahl.
Agentic Shopping Assistant
An Agentic Shopping Assistant combines language understanding, follow-up question logic and attribute-based matching. It works in multiple stages: capture needs, search the catalog, justify the result – rather than relying on static rules or scripts.
Not to be confused with:
FAQ chatbots with predefined responses. FAQ chatbots react to pre-formulated questions; an Agentic Shopping Assistant understands free formulations and adapts its strategy.
Practical application:
Increases consulting quality for large catalogs with many variants. Actively asks about budget and preferences before products are prioritized.
Guided Selling with AI
Guided Selling leads users step by step to the right selection – through questions, attribute-based filters and justified recommendations. AI prioritizes the most meaningful next questions based on the previous conversation.
Not to be confused with:
Static product filters without dialogue. Filters reduce the search space according to predefined criteria; Guided Selling with AI dynamically adapts the question flow to the situation.
Practical application:
Particularly effective for products requiring explanation or complex tariff logic. Entry via use case rather than technical specification.
Attribute-Based Product Matching
Attribute-based product matching checks requirements directly against concrete product attributes – not against text similarity. Must-criteria are explicitly validated; can-criteria flow into the ranking. The result is traceable and explainable.
Not to be confused with:
Reiner Vektorähnlichkeit ohne harte Attributprüfung. Vektorähnlichkeit misst semantische Nähe, prüft aber keine konkreten Werte wie „Akkulaufzeit ≥ 10h".
Practical application:
Reduziert Fehlkäufe, weil Muss-Kriterien explizit geprüft werden. Beispiel: „Akkulaufzeit ≥ 10h, Gewicht < 1,5 kg" als harte Filterkriterien vor dem Ranking.
Purchase Intent Recognition
Purchase intent recognition extracts the goal behind a search query: informational, comparative or purchase-ready. Systems that correctly classify intent can adjust priority and response strategy – e.g. comparison instead of direct purchase suggestion.
Not to be confused with:
Mere keyword extraction. Keywords describe what someone types; purchase intent recognition infers why someone is searching.
Practical application:
Bessere Priorisierung zwischen Beratung, Vergleich und Checkout-Nähe. Beispiel: „Beste Kamera fürs Reisen" → Vergleichsintention, kein spezifisches Modell.
Personalized Product Recommendation
Personalized product recommendations adapt results to context, captured needs and preferences. Personalization can be rule-based (e.g. after budget specification) or model-based (e.g. through usage patterns).
Not to be confused with:
Standardisierten „Kunden kauften auch"-Blöcken. Diese basieren auf Kaufhistorie, nicht auf dem individuellen Bedarf in der aktuellen Sitzung.
Practical application:
Higher relevance per session when preferences are cleanly captured. The recommendation varies depending on budget, use case and risk affinity.
Advisory Product Search
Beratende Produktsuche verbindet Suche und Beratung: verstehen, rückfragen, filtern, begründen. Sie beantwortet „Was passt zu meinem Bedarf?" statt nur „Was enthält den gesuchten Begriff?" – und leitet aus unklaren Eingaben schrittweise präzise Anforderungen ab.
Not to be confused with:
Traditional on-site search. On-site search returns results based on terms; advisory product search actively clarifies the need.
Practical application:
Particularly valuable for long decision cycles and high variant diversity. The system guides the user from a vague query to a concrete selection.
Expert Knowledge Training for Consulting
Expert knowledge training refers to the targeted adaptation of response style, prioritization rules and domain knowledge to customer specifications. Changes are versioned and made verifiable – without uncontrolled fine-tuning of the overall model.
Not to be confused with:
Uncontrolled fine-tuning without governance. Fine-tuning changes the base model globally; expert knowledge training is rule-based, versioned and reversible.
Practical application:
Kunden können festlegen, wie Antworten formuliert werden und welche Kriterien Vorrang haben. Beispiel: „Immer zuerst Sicherheit und Folgekosten erklären, dann Preis."
Compatibility Matching (Fitment Matching)
Compatibility matching checks whether a product is compatible with another system, model or component – based on hard technical attributes such as model number, series or year. AI systems with fitment matching only return results where all compatibility conditions are fully met.
Not to be confused with:
Textähnlichkeit oder Keyword-Matching. Ein Keyword-Match auf „Miele" findet alle Miele-Produkte; Fitment Matching prüft, ob der Artikel mit dem konkreten Modell kompatibel ist.
Practical application:
Beispiel: „Staubbeutel für Miele C2" → nur Staubbeutel, die als kompatibel mit Miele Complete C2 ausgezeichnet sind. Modell, Serie und Baujahr sind Muss-Kriterien.
Frequently Asked Questions about the AI Glossary
What is the difference between Agentic AI and a chatbot?
Chatbots follow scripts or answer predefined questions. Agentic AI plans autonomously, coordinates multiple sub-agents and performs multi-step tasks – without requiring human input at every step.
What does RAG stand for and when is it useful?
RAG stands for Retrieval-Augmented Generation. It is useful when a language model needs to access current or company-specific documents – such as policy texts in insurance or internal knowledge bases.
What is the difference between Vector Search and Agentic Search?
Vector Search finds similar documents based on embeddings. Agentic Search understands the intent of a query, asks follow-up questions when needed and matches product attributes directly – without a vector database.
What are AI hallucinations and how do you prevent them?
Hallucinations are factually incorrect statements that LLMs formulate with apparent certainty. Prevention: structured data sources instead of free generation and – for product recommendations – Agentic Search, which works directly on verified catalogue data.
What does GDPR-compliant AI mean in practice for online shops?
Personal data must be processed within the EU legal framework. Automated recommendations must be documented and contestable by customers. AGINITY AI meets these requirements through EU hosting and attribute-based, explainable recommendations.
What is Query Intent and why does it matter?
Query Intent is the actual purchase intent behind a search query. A system that understands intent finds matching products even with imprecise phrasing – and thus reduces zero results and cart abandonments.
What is a Zero-Result and how do you avoid it?
A Zero-Result occurs when the search returns no results despite products being available in the catalogue. Cause: exact keyword matching without synonyms or intent understanding. Agentic Search structurally eliminates zero results through attribute-based matching.
What is a POC at AGINITY AI?
A POC (Proof of Concept) at AGINITY AI is a 48-hour test with your real product data. You see live how Agentic Search answers customer queries and measure KPIs such as zero-result rate, conversion and session length.
How does AI check compatibility for spare and accessory parts?
Fitment Matching checks requirements such as model number, series and year of manufacture directly against hard product attributes – not text similarity. Only products that fulfil all compatibility conditions are returned as results.
Experience AgenticSearch Live
Test Agentic Search with your own product data – ready to go in 48 hours.
Further Reading
Author
Peter Niedermeier
Founder & CEO, AGINITY AI
Founder & CEO of AGINITY AI. Over 15 years of experience in e-commerce and AI product development. Develops Agentic Shopping solutions for European online retail.
View LinkedIn Profile