Retrieval-Augmented Generation RAG 📚 Look up relevant information first, paste it into the prompt, then let the model answer from it.
Grounding Making sure the answer is based on real source material rather than the model's impression of the world.
Hallucination confabulation 🌫 The model states something false with complete confidence, because plausible text is all it was ever optimised for.
Embeddings 🔢 Turning text into a list of numbers positioned so that similar meanings land near each other.
Vector Space latent space · embedding space The imaginary landscape where similar ideas sit close together and unrelated ones sit far apart.
Cosine Similarity A way of measuring how close two pieces of meaning are, ignoring how long the text was.
Semantic Search 🔎 Searching by what you meant rather than the exact words you typed.
Chunking ✂️ Cutting long documents into pieces small enough to search and to fit in a prompt.
Chunk Overlap ✂️ Letting each piece share a little text with its neighbours so sentences aren't cut in half.
Document Loading 📄 Getting text out of whatever format it was trapped in.
Vector Store vector database 🗄 A database built for finding the nearest meanings fast, even across millions of pieces.
Approximate Nearest Neighbour ANN · HNSW Finding almost-certainly the closest matches, very fast, instead of definitely the closest, slowly.
Indexing 🗄 The one-off job of preparing your documents so they can be searched quickly later.
Retrieval 🔎 The step that actually goes and fetches the passages most likely to answer the question.
Top-k 🔎 How many pieces of information you hand the model.
Recall & Precision Recall: did you find everything relevant? Precision: was everything you found relevant?
Reranking 🥇 Taking the passages search found and reading each one properly against the question to reorder them.
Cross-Encoder 🥇 A model that reads the question and the passage together, which is slower but much more accurate.
Bi-Encoder 🔢 A model that encodes questions and documents separately, so documents can be indexed ahead of time.
Hybrid Search 🧲 Running keyword search and meaning search together and merging the results.
BM25 keyword search · sparse retrieval 🧲 Classic keyword search: rank documents by which rare query words they contain.
Reciprocal Rank Fusion RRF 🧲 Merging two ranked lists by position rather than by score, since the two scores aren't comparable.
Query Rewriting ✏️ Turning what the user actually typed into something a search engine can work with.
HyDE hypothetical document embeddings ✏️ Have the model write a fake ideal answer, then search using that instead of the question.
Query Expansion ✏️ Adding synonyms and related terms to the search so you catch documents that used different words.
Multi-Hop Retrieval Questions that need two lookups — find A, then use A to find B.
Agentic RAG 🤖 Letting the model decide when to search again, what to search for, and when it has enough.
Citations attribution 🔗 Saying which source each claim came from, so a reader can check it.
Attribution 🔗 Tracing a statement back to the exact sentence that supports it.
Knowledge Graph GraphRAG Storing facts as things and the relationships between them, rather than as paragraphs.
Contextual Retrieval ✂️ Adding a sentence of surrounding context to each chunk before storing it, so isolated fragments still make sense.