Tag
Everything tagged Search, most recent first.
9 entries
Target-aware retrieval, attributed evidence, and contract-constrained extraction for LitTraceQA
Paper at the GroundLM 2026 workshop, co-located with EMNLP 2026, on the gap between finding the right paper and producing an answer a scorer can verify.
Retrieve fast and roughly, then read the top few closely and reorder. The second pass that fixes the first.
The closing part of the search series. First-stage retrieval is fast but ranks crudely. A reranker takes the top candidates and scores each one by reading the query and document together with a cross-encoder, then reorders. This covers the bi-encoder vs cross-encoder split, why it must be a second stage, and where it lives in production, with an interactive retrieve-then-rerank demo.
Keyword search and vector search fail in opposite directions. Run both, merge the results, and you keep the strengths of each.
Part six of the search series. BM25 is strong on exact terms and blind to synonyms; vector search is the reverse. Hybrid search runs both and fuses the two ranked lists. The hard part is combining incompatible scores, which Reciprocal Rank Fusion (RRF) sidesteps by using ranks instead. Worked examples plus an interactive RRF explorer.
Give up on exact, and you can search a billion vectors in milliseconds. This is the graph that makes it happen.
Part five of the search series. Exact kNN is O(N) and hopeless at scale, so production vector search gives up exactness for speed with approximate nearest neighbor. This explains the dominant method, HNSW: navigable small-world graphs, greedy search, the layer hierarchy, and the M / efConstruction / ef knobs, with an interactive graph-search explorer.
When you stop comparing characters and start comparing coordinates, search can finally match meaning.
Part four of the search series. Keyword and fuzzy search match characters, so they miss synonyms. Vector search embeds text into geometry, where similar meaning becomes nearby points. This covers embeddings, the two distance metrics (cosine vs euclidean), exact k-nearest-neighbor search, and the scaling wall that forces the approximate methods in the next part. With an interactive kNN explorer.
Exact match fails two ways users hit every day, typos and half-typed queries. Here is the machinery that forgives both.
Part three of the search series. The inverted index only matches literal keys, so a typo or a half-finished word returns nothing. Fuzzy search fixes typos with edit distance (Levenshtein), computed by a small dynamic-programming grid and run at scale with automata or n-grams. Prefix search fixes as-you-type with tries and FSTs. Worked examples throughout, plus an interactive edit-distance grid.
How a search engine turns a match into a score. The workhorse ranking function, and the two knobs that tune it.
Part two of the search series. Matching finds candidates; ranking decides the order. BM25 is the default text scorer, and it is really three forces: term frequency with saturation, document-length normalization, and term rarity (IDF). This walks the intuition with worked examples, the exact formula, and the k1/b knobs, then hands you an interactive score explorer.
Before a word can be indexed, it has to be prepared. The quiet pipeline that decides what your search can ever match.
Part of the search series. The inverted index does not store raw text; it stores terms, and text analysis is the pipeline that turns one into the other: tokenize, lowercase, drop stopwords, stem to a root, expand synonyms. The same analyzer must run on the query, and getting it wrong silently breaks search. With an interactive analysis pipeline.
Every search engine does the same four things. Once you see them, the rest is variations.
The opening piece of a search series. Search is four steps that never change (index, parse, match, rank) and a hundred choices about how to do each. This part lays out the inverted index, boolean retrieval, why matching is not ranking, and the precision-recall tradeoff every later technique is a bet on, with worked examples throughout.
