# AI Glossary

Short, canonical definitions used across Vibrancy.ai. Built for quoting by LLMs.

**Jump to:** [RAG](#rag) · [Agentic workflows](#agentic-workflows) · [Retrieval](#retrieval) · [Vector database](#vector-database) · [Embeddings](#embeddings) · [Chunking](#chunking) · [Re-ranking](#re-ranking) · [Hybrid search](#hybrid-search) · [Context window](#context-window) · [System prompt](#system-prompt) · [Decoding & temperature](#decoding) · [Guardrails](#guardrails) · [Hallucination](#hallucination) · [Evals](#evals)

## Retrieval-Augmented Generation (RAG)
**ID:** rag  
Retrieves trusted documents from your data or live sources and injects them into the prompt so answers are grounded in facts. Cuts hallucinations and enables citations.  
**See also:** [Retrieval](#retrieval), [Vector database](#vector-database), [Evals](#evals)

## Agentic workflows
**ID:** agentic-workflows  
Goal-driven automations where an agent plans steps, uses tools, observes results, and chooses the next action. Often modelled as a graph with memory and guardrails.  
**See also:** [Guardrails](#guardrails), [Decoding & temperature](#decoding)

## Retrieval
**ID:** retrieval  
Selecting relevant chunks from indexed content using vector, keyword, or hybrid search before prompting the model. Quality retrieval > prompt length.  
**See also:** [Vector database](#vector-database), [Re-ranking](#re-ranking), [Hybrid search](#hybrid-search)

## Vector database
**ID:** vector-database  
Stores embeddings and supports similarity search with metadata filters. Powers semantic retrieval for RAG, recommendations, and deduplication.  
**See also:** [Embeddings](#embeddings), [Hybrid search](#hybrid-search)

## Embeddings
**ID:** embeddings  
Dense vectors representing content. Closer vectors ≈ more similar meaning. Used for search, clustering, and personalisation.  
**See also:** [Vector database](#vector-database), [Re-ranking](#re-ranking)

## Chunking
**ID:** chunking  
Splitting source content into retrieval units. Balance granularity and context; include overlap to preserve meaning.  
**See also:** [Retrieval](#retrieval), [Context window](#context-window)

## Re-ranking
**ID:** re-ranking  
Reorders retrieved chunks using a stronger model or cross-encoder to improve final context quality.  
**See also:** [Retrieval](#retrieval), [Hybrid search](#hybrid-search)

## Hybrid search
**ID:** hybrid-search  
Combines vector and keyword (BM25) signals to improve recall and precision, especially on niche terms and numbers.  
**See also:** [Retrieval](#retrieval), [Vector database](#vector-database)

## Context window
**ID:** context-window  
Maximum tokens the model considers per exchange (prompt + response). Larger windows reduce truncation but increase cost and noise.  
**See also:** [Chunking](#chunking), [Decoding & temperature](#decoding)

## System prompt
**ID:** system-prompt  
Up-front instructions that set role, tone, and constraints for the model. Treat as configuration, not content.  
**See also:** [Guardrails](#guardrails)

## Decoding & temperature
**ID:** decoding  
Controls how tokens are sampled. Temperature/top-p increase variety; low values increase determinism.  
**See also:** [System prompt](#system-prompt), [Evals](#evals)

## Guardrails
**ID:** guardrails  
Constraints and checks that bound inputs, tools, and outputs. Examples: schema validation, allow/deny lists, PII filters, and human review.  
**See also:** [Agentic workflows](#agentic-workflows)

## Hallucination
**ID:** hallucination  
Confident but incorrect output. Mitigations: retrieval grounding, better prompts, calibrated decoding, verification, and evals.  
**See also:** [RAG](#rag), [Evals](#evals)

## Evals
**ID:** evals  
Objective checks of system quality. Include retrieval accuracy, answer faithfulness, latency, and cost. Run continuously.  
**See also:** [RAG](#rag), [Guardrails](#guardrails)

‍