From the inverted index to agents that search for themselves.
Indexes, postings lists, and Boolean AND.
A count in the note, times how rare the word is.
TF-IDF with a cap on repeats and a fair length penalty.
Index, BM25, and a top-k heap in one file.
Words as points, learned from the company they keep.
One vector per sentence, trained with contrastive pairs.
Two encoders, one dot product, search by meaning.
Approximate nearest neighbours: IVF, HNSW, and filters.
Retrieve wide with a bi-encoder, reorder with a cross-encoder.
Fuse BM25 and dense ranks with reciprocal rank fusion.
Fix, expand, and route the query before you search.
Retrieve, pack the context, then generate with citations.
How you split the text decides what you can find.
Grade the retrieval, retry, and know when not to retrieve.
Recall@k, MRR, nDCG, and grading the answer.
Entities and edges for questions that span documents.
One space for images and text, and searching page images.
Stages, budgets, caches, and freshness at scale.
A model that plans, searches, reads, and searches again.
Case study: a search agent trained to retrieve.