The AI, RAG & website chatbot glossary.
Clear, approachable definitions of the technical and operational concepts behind modern AI website assistants. Every term is engineered as a factual citation anchor.
Retrieval-Augmented Generation (RAG)
Retrieval-Augmented Generation (RAG) is an AI architecture that first retrieves relevant text passages from a private knowledge source, then feeds those passages into a language model to generate an answer grounded strictly in verified facts.
Vector Embeddings
A vector embedding is a high-dimensional numerical representation of a piece of text where semantically similar concepts are placed close together in mathematical vector space.
Semantic Chunking
Semantic chunking is the practice of dividing a long document or web page into discrete, logically coherent segments while preserving contextual meaning and structural boundaries.
Source Grounding & Citations
Source grounding is the enforcement mechanism ensuring an AI model generates responses derived strictly from retrieved reference documents, accompanied by deep-link citations back to the source material.
Context Window Limits
A context window is the memory capacity of a large language model, defined as the total token limit it can simultaneously ingest across system prompts, retrieved knowledge chunks, and dialogue history.
Cross-Encoder Reranking
Cross-encoder reranking is a secondary relevance scoring pass that jointly evaluates the query and retrieved candidate chunks to reorder them with high semantic precision before LLM synthesis.
Multi-Domain Knowledge Pooling
Multi-domain knowledge pooling is the aggregation of multiple distinct domains, subdomains, and external document repositories into a unified, cross-referenced vector knowledge base.
Live Chunk Editing
Live chunk editing is the capability to directly alter the textual content and vector embedding of a specific knowledge segment in real time without triggering a full website recrawl.
Vector Database & pgvector
A vector database is a specialized database system built to store, index, and query high-dimensional vector embeddings using spatial similarity algorithms such as HNSW or IVFFlat.
Dense Semantic Search
Dense semantic search is an information retrieval technique that matches queries to documents based on their conceptual meaning in vector space rather than literal keyword overlap.
Hybrid Search (Dense + Sparse Retrieval)
Hybrid search is a multi-modal retrieval architecture that combines dense vector embeddings with sparse lexical search (e.g. BM25/tsvector) using rank fusion algorithms.
Reciprocal Rank Fusion (RRF)
Reciprocal Rank Fusion (RRF) is an algorithm that combines the ranked results of multiple search algorithms (such as vector and keyword search) into a single unified ranking based on their reciprocal positions.
Cosine Similarity Thresholds
Cosine similarity is a mathematical metric that measures the cosine of the angle between two non-zero vectors in inner product space, outputting a similarity score between -1.0 and +1.0.
Full-Text tsvector Indexing
A tsvector is a sorted list of distinct normalized lexemes in PostgreSQL, optimized for linguistic stemming and high-speed full-text keyword retrieval.
Trigram Fuzzy Matching (pg_trgm)
Trigram search is a substring indexing technique that decomposes words into three-character tuples to compute similarity and match queries despite spelling mistakes and typos.
AI Hallucination
An AI hallucination is a phenomenon where a large language model generates factually incorrect or fabricated statements presented with high linguistic confidence, unsupported by real-world context.
Prompt Injection & Jailbreaking
Prompt injection is a security vulnerability where an attacker embeds adversarial instructions into user input to force an LLM to disregard its original system constraints, reveal hidden prompts, or execute unauthorized commands.
LLM Output Guardrails & Validation
LLM guardrails are programmatic pre- and post-processing validation layers designed to enforce domain boundaries, block toxic or adversarial content, and prevent data leakage.
Enterprise Zero-Data Retention
Zero-data retention (ZDR) is an enterprise privacy architecture ensuring that customer data, conversational queries, and uploaded documents are never retained by third-party AI providers for model training.
Multi-Tenant Workspace Isolation
Multi-tenant workspace isolation is a database and architectural security pattern that enforces strict data separation between different customer workspaces across all database queries.
System Prompt Architecture
A system prompt is a high-priority directive passed to an LLM before conversation turns, establishing the assistant’s persona, boundary rules, and grounding constraints.
Time-To-First-Token (TTFT)
Time-To-First-Token (TTFT) is a performance metric measuring the latency from initial user request dispatch to the arrival and display of the very first generated token.
Server-Sent Events (SSE Streaming)
Server-Sent Events (SSE) is a unidirectional HTTP streaming standard that allows a server to push real-time text streams to web clients over a persistent connection.
Shadow DOM CSS Isolation
Shadow DOM is a browser web standard that provides scoped encapsulation for a component’s DOM tree and styles, completely isolating it from the main document’s CSS rules.
Function Calling (AI Tool Use)
Function Calling is an architecture where an AI model identifies user intent requiring external actions and generates structured JSON parameters to execute designated webhook tools.
Outbound Webhooks
Outbound Webhooks are automated HTTP POST notifications dispatched to external endpoints when specific events—such as captured leads or answered inquiries—occur.
Website Crawler Concurrency
Crawler concurrency is the rate at which an automated web crawler dispatches parallel HTTP requests to discover, extract, and index multiple URLs simultaneously.
Sitemap XML Ingestion
Sitemap XML ingestion is the automated parsing of a website’s standard XML sitemap file to discover all canonical URLs, priority tiers, and last-modified dates for indexing.
Headless Browser vs Cheerio Crawling
The extraction distinction between fast HTTP HTML parsing (Cheerio) and headless browser automation (Playwright) capable of executing client-side JavaScript before content extraction.
Support Ticket Deflection Rate
Ticket deflection rate is the percentage of customer questions and support requests resolved autonomously by self-serve AI without escalating to human support agents.
First-Response Time (FRT)
First-Response Time (FRT) is a customer service key performance indicator measuring the elapsed time between initial ticket creation and the first agent reply.
Support-Driven Churn Deflection
Support-driven churn deflection is the reduction in customer turnover achieved by eliminating support wait times and resolving pre-churn blockers instantaneously.
Self-Serve Customer Support
Self-serve customer support is a support architecture that enables customers to autonomously access information and resolve problems without human intervention.
Competitor Model Multiplier Penalty
A model multiplier is a pricing penalty imposed by chatbot vendors where using standard LLMs (like GPT-4o) consumes 3 to 6 credits per single customer message.
White-Label AI Chatbots
White-label AI is software that allows businesses and marketing agencies to completely remove vendor logos and branding, deploying the tool to clients as their own proprietary platform.
Conversational Lead Capture
Conversational lead capture is the automated collection and qualification of visitor contact details inside an AI chat dialog, passing leads directly to CRMs or webhooks.
Generative Engine Optimization (GEO)
Generative Engine Optimization (GEO) is the practice of optimizing digital content for discovery, comprehension, and citation by generative AI search engines and answer bots.
Answer Engine Optimization (AEO)
Answer Engine Optimization (AEO) is a sub-discipline of SEO focused on formatting content into concise, direct answers optimized for search engine answer boxes and AI voice assistants.
Test it on your own website in under 2 minutes.
Enter your domain to index your pages and preview live answers.