Glossary

The terms that show up in the challenges, in the order you meet each one across the worlds.

Welcome Village

AI (Artificial Intelligence): Systems that make predictions or generate content from data — without hand-written rules for every case.
LLM (Large Language Model): A model trained on lots of text that predicts the most likely next word — that's how it "talks".
Prompt: The text you send the model. How it's written changes the quality of the answer a lot.
Token: A chunk of text (a word or part of one) the model reads and generates. Cost and context limits are measured in tokens.
Context: Everything sent along with the prompt (history, documents, instructions) so the model can answer more precisely.
Hallucination: When the model confidently makes up an answer with no real basis — the main reason to double-check its output.

Few-shot: Giving 2–3 examples in the prompt so the model follows the pattern, instead of just explaining the task.
System prompt: The instruction "behind" the conversation that sets the model's persona and rules before any user question.
Prompt injection: Trying to manipulate the model with malicious text hidden in the input so it ignores its original instructions.

Training vs. Inference: Training is when the model learns from data; inference is the already-trained model answering something new.
Generalization: The model's ability to get examples it never saw right — what separates real learning from memorizing.
Latency: The time between asking and getting the answer. Along with cost, it limits how much processing you can afford.

RAG (Retrieval-Augmented Generation): Fetching relevant information before generating the answer, instead of relying only on what the model "memorized" in training.
Keyword search: Search that matches exact terms in the text — breaks on synonyms ("car" won't find "automobile").
Embedding: A text represented as a vector of numbers, where texts with similar meaning end up "close" together.
Cosine similarity: A metric measuring how much two embeddings point in the same direction — used to find the chunk closest to the question.
Chunk / Chunking: Splitting a document into smaller pieces before indexing. Big chunks dilute the signal; small ones lose context.

RAG pipeline: The retrieve → build prompt → generate answer sequence — the skeleton of any RAG system.
Citation (grounding): Tracing each claim in the answer back to the source passage that backs it — without it, hallucination slips through.
Hybrid search: Combining keyword search with vector search, tuning the weight of each for the situation.

Reranking: Reordering an initial search's results with a more expensive, precise model, to push the best ones to the top.
Neighbor expansion: Pulling in the chunks around a matched result too, so context isn't lost at the chunk boundary.
Semantic compression: Indexing a summary of the text instead of the original, to cut noise — at the risk of losing fine detail.

Cache: Storing the result of a search or generation already made, to avoid paying (time and cost) for the same question twice.
Observability: Instrumenting the system to see what it's doing in production — latency, cost, error rate per step.
Ingestion: The pipeline that takes raw documents (e.g. PDF), cleans, chunks and indexes them — before any search can happen.
Vector database: A database built to store and search embeddings quickly at scale.

Knowledge graph: Entities and relations extracted from text, represented as nodes and edges — instead of loose text.
GraphRAG: RAG that traverses a knowledge graph instead of (or alongside) vector similarity search.
Multi-hop query: A question only answerable by chaining several facts together — usually needs graph traversal, not a single chunk.

Terminology consistency: Making sure the same concept is always called by the same term — inconsistency confuses both the user and the retriever.
Concept drift: When a term's meaning shifts over time in the knowledge base, and the system keeps using the old definition.
Confidence gate: A rule deciding whether the system answers or admits uncertainty, based on how reliable the retrieval was.