Discover more

AI

• Agentic AI: autonomous agents based on LLMs that can make decisions and execute complex tasks.
• Boost / Bury: promotion / demotion of products based on explicit criteria in post-processing, within the Vertex AI Search ecosystem.
• Business rules / Safety rules / Deterministic rules / Guardrails: strict commercial rules applied in IT systems, including to AI results. E.g., stock, margin, compatibility, brand, compliance.
• Candidate Generation / Retrieval: initial stage of rapid selection of a subset of relevant results from a large catalog.
• Collaborative filtering: recommendations based on patterns in purchase history (those who bought X also bought Y).
• Cold Start: a situation where there is insufficient history for a new product or customer.
• Composite AI: combining multiple AI techniques (models + rules) to solve complex problems.
• Content-based filtering: recommendations based on product attributes/content (description, technical sheet, image).
• Deep Learning: models based on deep neural networks.
• Diversity: mechanisms that avoid overly repetitive recommendations and allow for "testing" of new options.
• Drift: degradation of model performance over time due to changes in data/behavior.
• Explainability / xAI / Feature attribution: the ability to justify and present the behavior of AI technologies to humans.
• Feature: input variable in an AI model. E.g., price, brand, time, device, last clicks.
• Feature Store: component for storing and consistently serving features to online or offline models.
• Feedback Loop: mechanism through which AI user behavior improves AI models. It can be positive (strengthens a detected association) or negative (weakens it).
• Filter: exclusion of products based on explicit criteria during the candidate generation phase, within Vertex AI Search for commerce.
• Funnel architecture: AI architecture for processing in successive stages (Retrieval, Ranking, Re-ranking), where the number of products analyzed decreases as precision increases.
• Generative AI / LLM: large language model, Deep Learning trained on a vast portion of human written knowledge, capable of generating text, images, etc. E.g., ChatGPT.
• Generative Configuration: automatic generation of technical configurations and BOMs using generative AI.
• GraphRAG: technology that combines vector search with knowledge graphs to reduce AI hallucinations regarding relationships and compatibility.
• Hybrid architecture: combination of probabilistic AI (models) and deterministic rules (guardrails), plus integration with existing systems (ERP, WMS, CRM).
• Lexical Search: search based on exact or approximate matching (typo correction) of words or phrases in text, without understanding the meaning.
• Popularity bias: the tendency to over-recommend best-sellers at the expense of the long-tail.
• Post-processing / Re-ranking: the stage where generated AI candidates are adjusted or filtered by score using rules.
• Privacy-first architecture: AI integration design that prioritizes company data protection and compliance.
• Ranking / Scoring: ordering generated AI candidates based on a score (e.g., probability of conversion) and post-processing.
• Semantic Search: the goal of AI technologies to understand human language; primarily uses vector search along with other specialized components.
• Two-tower architecture: AI architecture characteristic of Deep Learning with two representations (e.g., product and customer context) compared in a vector space.
• Vector / Embedding: a vector is a mathematical structure (a list of numbers); an embedding is a specific vector generated by an AI model that captures the meaning of a text or product/customer.
• Vector Search: search for the nearest "neighbors" in embeddings space (by numerical similarity across vector dimensions).

Business

• AOV (Average Order Value): The average value of an order.
• BOM (Bill of Materials): structured list of components required for a product or an offer.
• Churn Rate: measures customers who stop buying out of the total customer base.
• CLV (Customer Lifetime Value): total estimated value generated by a customer over the duration of their relationship with the company.
• CTR (Click-through rate): the percentage of clicks on recommendations out of total impressions.
• Long tail: the situation where there are many niche products (low volume) but with good margins and specific relevance.
• pCVR (predicted Conversion Rate): estimated probability of a conversion. E.g., purchase/acceptance of an offer.

Software

• A/B: controlled experiment (variant A vs. B) to compare KPIs between versions.
• Backtesting: offline evaluation on historical data before launch.
• Blueprint: implementation plan that guides execution (e.g., software).
• Build / Custom: software application developed "from scratch," usually with control over the code.
• Cache invalidation: deleting data from cache when source system data changes. E.g., stock/price, so as not to serve stale results.
• Clickstream: the sequence of user interactions (views, clicks, basket).
• Data contract: technical agreement for integration between systems regarding fields, types, meanings, validations, and responsibilities for the data transmitted between them.
• Data Warehouse: centralized data repository for analysis and training AI models.
• Fallback: backup strategy when an integration, including AI, has no result or the API does not respond.
• GIGO (Garbage In, Garbage Out): if data is wrong or incomplete, integration results will be poor or non-existent, regardless of technology.
• Latency: system response time (important for UI recommendations).
• Managed / Cloud Native: software application paid for by consumption to vendors (usually cloud) where you can integrate and configure company processes and data without having access to the code.
• Placement: the location in the UI where recommendations appear. E.g., homepage, product page, cart, quoting.
• Shadow mode: running a technology internally in parallel with the current system to compare on real flows if improvements exist, without displaying them to users.
• SLA / SLO: availability and performance commitments (SLA) or targets (SLO).

Guides and Technical Articles

Web Pages

Talk with the author for a free technical audit.