SHEN YIFull-Stack & AI

Paris • Singapore • Shanghai

RAG Formation · Reference

Terminology glossary — EN

A standalone cheat sheet for learning and testing RAG quality on this site.

$ reference --status
▸ course RAG Formation
▸ mode standalone reference
▸ status Complete
● learn → test → verify

Reference sheet

Terminology glossary — EN

The guided RAG Formation course content is complete; this reference is available for self-directed practice.

Why this exists

Terminology drift compounds: once a lesson translates "grounding" one way, every later lesson that reuses the concept either repeats the mistake or introduces a second, inconsistent translation. This page is the single source every lesson, cheatsheet, and animation must match. It is a living document — extend it before writing new localized content that introduces a term not yet listed here, rather than inventing wording ad hoc.

Locked terms

EnglishDefinition
Chunking Splitting a source document into retrievable windows, with overlap to avoid cutting a fact away from its identifying context (see Lesson 6).
Reranking Re-sorting a retriever's coarse candidate list using harder signals — exact-ID hits, metadata, cross-encoder scores (see Lesson 8).
Hallucination The system answers a question the corpus does not support, instead of refusing or saying it doesn't know (see Lesson 12).
Grounding / grounded An answer is grounded when every claim in it is backed by retrieved evidence the model actually saw (see Lesson 1).
Refusal The system declining to answer because the corpus lacks sufficient evidence, rather than guessing (see Lesson 3, Lesson 12).

Scope and status

  • Sourced from current usage already committed across lessons 1, 3, 6, 8, and 12 — this page locks existing wording, it does not introduce new translations.
  • English only. French is paused course-wide (see rag-formation/CLAUDE.md — "Sequencing"); a French column will be added when French rollout resumes.
  • Five terms today because these are the ones flagged as highest drift-risk (reused across many lessons, and now also baked into animation captions per animations/ANIMATIONS-PLAN.md). Extend this table — don't start a second glossary file — when another term shows the same reuse pattern.

Continue learning: Open the matching standalone lesson · Course roadmap