Glossary

One-line, plain-English definitions for the terms used across these notes. Skim it before an interview, or jump here whenever a word is unfamiliar.

Core LLM terms

LLM (Large Language Model) — a model that predicts the next token over and over; effectively a function from text to a probability distribution over the next token.
Token — the unit a model reads; a word-piece. ~1 token ≈ 4 characters. Cost and context limits are counted in tokens.
Context window — the max tokens a model can consider at once (prompt + output).
Temperature — randomness of sampling; low = focused/deterministic, high = creative.
Top-p / top-k — limit which tokens are eligible before sampling (nucleus / top-k).
Hallucination — a confident but unsupported/false answer the model invents.
Reasoning model — a model trained to "think" (long internal chain) before answering; better at multi-step problems, slower and pricier.
System / user / assistant — message roles; system = standing rules, user = the request, assistant = the model's prior turns.

RAG (Retrieval-Augmented Generation) — retrieve relevant text at query time and put it in the prompt so the model answers from it.
Indexing — the offline phase: load → chunk → embed → store.
Querying — the live phase: embed question → retrieve → augment prompt → generate.
Document — LangChain's common unit: page_content (text) + metadata.
Loader — reads a source (PDF, CSV, SQL, …) into Documents.
Chunk — a small, independently-retrievable piece of a document.
Chunk overlap — shared text between neighbouring chunks so boundary facts survive.
Recursive splitting — split on paragraph → line → sentence → word; the default.
Grounding — tying the answer to retrieved source text (so it's citable).

Embedding — a fixed-length vector that captures the meaning of text.
Dimension — how many numbers in the vector (e.g. MiniLM = 384, OpenAI small = 1536).
embed_query — embed one string (the question).
embed_documents — embed a list of strings (your chunks) in one batch.
Cosine similarity — angle between two vectors; 1 = same meaning, 0 = unrelated.
L2 (Euclidean) distance — straight-line distance; lower = more similar (ChromaDB's default — opposite direction to cosine).
Normalization — scaling a vector to length 1 so cosine == dot product (faster).

Vector store — lightweight library for storing + KNN-searching vectors (FAISS, Chroma); great under ~1M vectors.
Vector database — full system with filters, CRUD, sharding, HA for production scale (Pinecone, Qdrant, Weaviate, Milvus).
Retriever — embeds a query and returns the closest chunks (as_retriever).
top-k — how many chunks retrieval returns.
Similarity search — similarity_search (plain) / similarity_search_with_score.
HNSW — a graph index for fast approximate nearest-neighbour search; the common default. IVF — a cluster-based ANN index.

LangChain — framework for composing LLM apps (loaders, splitters, chains).
LCEL (LangChain Expression Language) — compose a pipeline with the | operator.
create_stuff_documents_chain — "stuffs" retrieved docs into the prompt's {context} slot.
create_retrieval_chain — wires a retriever to a document chain = the RAG pipeline.
History-aware retriever — rewrites a follow-up into a standalone query using chat history before retrieving.

Hybrid search — combine dense (vector) + sparse (BM25/keyword) retrieval.
Reranking — a cross-encoder re-scores the shortlist for precision.
RRF (Reciprocal Rank Fusion) — merge ranked lists by rank, not score.
HyDE — embed a hypothetical answer instead of the question to retrieve better.
Multi-query — generate several phrasings, retrieve for each, merge.
Corrective RAG (CRAG) / Self-RAG / Adaptive RAG — agentic variants that check, correct, or decide when to retrieve.
Multimodal RAG — retrieval over text and images.
Guardrails — input/output filtering and injection defense.