LLM API
A hosted service that lets your app send prompts to a language model and receive generated text or structured output.
Compare LLM APIsAI definitions
Plain-English AI Terms for Builders and Students.
Use this glossary when a tutorial, FAQ, or provider page assumes vocabulary you have not learned yet. Each definition links back to a practical FreeGrid guide.
Last updated September 17, 2026. Free tiers, model access, and dataset terms can change; verify official sources before launch.
Embedding
A vector representation that lets software compare meaning, not only exact keywords.
Compare embeddingsRAG
Retrieval-augmented generation: retrieve relevant source chunks first, then generate an answer grounded in those chunks.
Build RAGVector database
A storage and search system for embeddings, metadata, and similarity queries.
Compare vector DBsReranker
A model that reorders retrieved chunks so the answer model sees the most relevant evidence first.
RAG model guideContext window
The amount of text or tokens a model can consider at once. Larger context helps, but better retrieval still matters.
Use context wiselyToken
A piece of text used for model input and output accounting. Token count affects cost, latency, and context limits.
Estimate costRate limit
A provider rule that caps requests, tokens, or usage over time. Good apps handle limits with retries and fallback UX.
Rate-limit guideAI agent
A workflow where a model plans or decides steps and calls tools under constraints, logs, and approvals.
Agent stackOCR
Optical character recognition extracts text from scanned documents or images before search, RAG, or analysis.
OCR alternativesFine-tuning
Training a model further on task-specific examples. It is useful sometimes, but many projects should improve retrieval and prompts first.
Model selectionLocal LLM
A language model running on your own computer or server instead of a hosted API.
Local vs APIMCP (Model Context Protocol)
An open standard that connects AI applications to tools and data - like a USB-C port for AI apps. Build one MCP server and every compatible client can use it.
MCP guideDiffusion LLM (dLLM)
A model that generates and refines many tokens in parallel instead of sequentially, trading the classic autoregressive pipeline for extreme speed - Mercury 2.5 exceeds 1,100 tokens/sec this way.
See it in the model waveSparse MoE
Mixture-of-Experts architecture where only a fraction of the parameters activate per token - DeepSeek V4.1 Flash runs 552B parameters while activating just 8B per input token, keeping costs near free-tier levels.
Compare cheap modelsCoordination tax
The token waste from running too many parallel sub-agents: they duplicate work and double-check each other without quality gains. OpenAI's Codex team recommends keeping it to about two.
Agent cost disciplineGEO (Generative Engine Optimization)
Making content quotable by AI search: direct answers first, FAQPage schema, dated freshness signals, and comparison tables that AI Overviews and answer engines can cite verbatim.
The GEO playbookStealth model
A model served through a router under an anonymous provider name - Union Alpha on OpenRouter is a free frontier-level example. Great for experiments; never send sensitive data.
Free model pathsStudy path
Learn the terms in build order.
Start with API, model, token, and latency. Then learn embeddings, vector databases, RAG, reranking, and evaluation. After that, compare local deployment, fine-tuning, observability, and cost controls.
Recommended order
- LLM API, prompt, token, latency
- Dataset, license, schema, citation
- Embedding, vector database, semantic search
- RAG, reranking, citation, evaluation
- Local model, fallback, observability
People also ask
Glossary questions.
What AI term should beginners learn first?
Start with API, model, prompt, token, and latency. Those terms appear in almost every AI project decision.
What is the difference between search and RAG?
Search retrieves relevant information. RAG uses retrieved information as context so a model can generate an answer with sources.
What is the difference between embeddings and fine-tuning?
Embeddings help retrieve similar information. Fine-tuning changes model behavior with training examples. Many RAG projects need better embeddings before they need fine-tuning.