Every term this site uses, in one place. Each card opens the same drill-down as the underlined phrases in the text: one line, what it is, what it is not, and how it shows up in the insurance example.
Two sentences hold the words apart. Retrieval is a query; the model reads only what it returns. And: the model recommends; a person decides.
AbstentionThe writer declines to recommend and hands the case to a person, because no relevant current passage scored high enough or an applicable rule's passage was not retrieved.ACL filterA condition in the retrieval query that restricts the search to rows the caller may see; ACL stands for access control list.Agent RetrievalA managed search engine that stores Collections of JSON Data Objects with named vectors and runs vector, text, semantic and hybrid search over them.Agent SearchA product for building enterprise search and recommendation apps over data stores, with its own ranking and generative answers.Agentic retrievalInstead of one retrieval before generation, the model chooses when to retrieve and with what query, reads the results, and may call again.AlloyDB for PostgreSQLA fully managed, PostgreSQL-compatible database on Google Cloud that offers pgvector and its own ScaNN vector index.ANN indexAn approximate nearest neighbor index: a data structure built over the vectors in advance so that a query compares only a small share of the rows, at the risk of missing some true neighbors.Answer correctnessWhether the output matches the expected answer: recommendation, citations, evidence ids and required fields.Applicant similarityThe retrieval question of Corpus A: which past submissions look like this one, by occupancy, construction, size, age and exposure.AuthorityThe rank of a source in deciding a recommendation: current guidance decides, superseded guidance and precedents only inform, comparables decide nothing.Auxiliary vector tableA table that holds one vector per source row together with that row's id; searches rank its rows and join back to the source table.BackfillA job that computes the new value, here a new vector, for every existing row, in batches, while the system keeps serving.Benchmark harnessA program that runs the same query set against each configuration, computes the answer key with exact search, and records recall, work, short results and latency per setting in one table.Bi-encoderAn encoder that turns the query and each document separately into vectors, which are then compared with a distance.BM25A ranking function for keyword search that scores a document by term frequency (with saturation), term rarity across the corpus, and document length.Bound parameterA placeholder in a SQL statement ($1 in PostgreSQL, ? in SQLite) whose value is supplied at execution time, so the value never becomes part of the SQL.Btree indexA balanced, sorted tree over a column that lets the database jump to matching values instead of reading every row.Built-in embeddingA feature where you send raw text and the store calls an embedding model at write time and at query time, so you never handle the vectors.Candidate poolThe first stage's output that the reranker scores: typically 50 or 100 rows per query; any relevant row outside it cannot be recovered by reranking.Candidate setThe top k rows of the first-stage ranking (dense, lexical or both), 50 by default on this site, that fusion and reranking work on.ChromaAn Apache 2.0 store for embeddings with metadata, used embedded on a laptop, self-hosted, or as Chroma Cloud.Chunk overlapA number of characters or tokens shared by neighboring chunks, so a sentence near a cut appears whole in at least one of them.ChunkA piece of a document, with its own vector and metadata, that retrieval can return on its own.ChunkingThe ingestion step that cuts parsed documents into chunks, by size, by structure or by meaning.CitationAn id in the answer (rule id, passage id, evidence id) that a reviewer or a program can look up and compare with what retrieval returned.Cloud SQL for PostgreSQLA fully managed relational database service on Google Cloud; its PostgreSQL edition offers the pgvector extension for vector columns.Content hashA digest such as SHA-256 computed from the exact bytes of a text: the same text gives the same hash, any change gives a different one.Context budgetThe per-query limit on how many retrieved passages the select stage hands to the answer writer: 5 on this site.Context windowThe text a model reads for one request, measured in tokens and limited in size by the model.CorpusThe collection of stored units, rows, documents or chunks, that one retrieval searches.Cosine distanceOne minus the cosine similarity of two vectors, so that smaller means closer and an ascending sort returns the nearest rows.Cosine similarityThe dot product divided by the product of the two vectors' lengths: 1 for the same direction, 0 for perpendicular, -1 for opposite.Cross-encoderA model that takes a query and one passage as a single input and outputs one relevance score for the pair; it produces no embedding, so nothing can be precomputed.CutoverOne switch that moves queries to the new column, table, index or service, after the new one is complete and has passed evaluation.Data residencyThe requirement, or guarantee, that data stays in a named place: a machine, a region or a jurisdiction, at rest and while it is processed.Decision guideThe recipes page's table from a question to the chapter that answers it, plus a flow for choosing where vectors live.Dense retrievalRetrieval that embeds the query and every row as short vectors in which every slot holds a value, and ranks rows by vector distance; it connects paraphrases and blurs exact identifiers.DimensionOne position in a vector, and by extension the vector's length: [0.9, 0.8] has dimension 2.Distance operatorAn operator such as <=> in pgvector that takes two vectors and returns a number, smaller meaning closer, so ORDER BY can sort rows by it.Document versionAn identifier of one edition of a source, stored on every chunk cut from it, so retrieval can filter to the current edition or explain an old one.Domain adaptationFurther training of an embedding model on pairs of related texts from your own domain, so that your vocabulary lands in sensible places.Dot productMultiply two vectors position by position and add the results; larger means more aligned, and longer vectors score higher.Dual writeWriting every new or changed row to both the old and the new structure, so the new one stays complete while it is being filled.ef_constructionHow many candidate neighbors HNSW keeps while inserting each point; a wider beam finds better neighbors and makes a better graph, at the cost of build time and insert speed.ef_searchHow many candidates an HNSW search keeps on the bottom layer; wider finds more true neighbors and computes more distances.ElasticsearchA search engine built for full-text search that also stores dense vectors and fuses keyword and vector results.Embedding benchmarkA fixed set of public datasets and tasks on which many embedding models are scored the same way, such as MTEB, the Massive Text Embedding Benchmark.Embedding migrationRecomputing every stored vector with a new embedding model, with the old vectors serving queries until the new ones are complete, evaluated and chosen.Embedding modelA trained model that reads text through its tokenizer and encoder, pools and usually normalizes, and writes a vector of fixed dimension.Embedding spaceThe space defined by one embedding model, in which texts with related meaning sit close together.EmbeddingThe fixed-length list of numbers an embedding model writes for one text, placing that text as a point in the model's space.Euclidean distanceThe square root of the summed squared differences between two vectors; smaller means closer.Evidence idA stable id for one piece of evidence, such as a provider's roof age, so a reason can say where a number came from.Exact identifierA token whose meaning comes from convention, not from its letters, such as a rule id, a submission id, an evidence id or a parcel id; only an exact match is a correct match.Exact searchA vector query that computes the distance from the query vector to every candidate row, sorts, and keeps the nearest k, so the result is the true top k.EXPLAIN ANALYZEA PostgreSQL command that executes a statement and prints its plan with the rows each step actually produced.FAISSA C++ library with Python wrappers for similarity search and clustering of dense vectors; code you call in your own process, not a database.FaithfulnessWhether an answer stays within what its context supports, judged claim by claim.Feature rerankerA reranker without a model: it combines facts each candidate already carries (exact band matches, recency, the first stage's similarity) into one score with configured weights.Field weightingScoring the fields of a row separately (title, id, body) and combining them with weights, so a query term in the title counts more than the same term in the body.FreshnessWhether a row reflects the current state: the current policy version, an active status, a recent decision.Full-text searchSearch that splits text into normalized words and finds documents containing the query's words through an inverted index.Gemini Enterprise Agent PlatformThe Google Cloud platform name under which Vector Search, Agent Retrieval and RAG Engine are documented since April 22, 2026.Graph RAGEntities and their relations are stored as a graph; retrieval follows the relations and hands the model the connected facts.Ground-truth searchThe exact top k for a query, filter and metric, computed by comparing every candidate, used as the reference to measure what a faster method missed.Grounded generationWriting an answer from the retrieved passages only, with each claim pointing to the passage or evidence id that supports it.halfvecA vector column type that stores each number in 2 bytes instead of 4, halving storage at a small cost in precision.HallucinationAn answer's claim that cannot be traced to anything retrieval returned, such as an invented limit or a citation to a rule that was not in the context.Hit rate@kPer query, 1 if any relevant row is in the top k, else 0; averaged over queries.HNSWAn approximate nearest-neighbor index made of layered proximity graphs: a search descends greedily through sparse upper layers, then runs a beam search on the bottom layer, which holds every point.Hybrid searchRunning a keyword (sparse) retriever and a vector (dense) retriever on the same query and filters, then merging the two ranked lists into one candidate list.HyDEA language model writes a hypothetical document that answers the question; its embedding is used as the query vector to retrieve real documents near it.Index recallOf the rows exact search ranks in the top k, the share an approximate index also returned, averaged over a query set.Instruction prefixA short instruction prepended to the query (and sometimes to documents) because the model was trained to see it there.Inverted indexAn index that maps each term to the documents containing it, like the index at the back of a book.Iterative scanA pgvector setting that lets an approximate index scan keep going when a filter removed too many rows, until enough rows pass or a limit is reached.IVFFlatAn approximate index that groups vectors into lists with k-means and, at query time, exactly scans only the lists whose centers are nearest the query.Keyword searchFinding texts that contain the query's words and ranking them by how strongly they contain them.LangChainA framework with model clients, document loaders, splitters and vector store integrations; its PostgreSQL integration is the package langchain-postgres.Late interactionQuery and document are encoded separately into one vector per token; at query time each query token's closest match among the document's tokens (MaxSim) is found and the maxima are summed.LlamaIndexA framework that builds a VectorStoreIndex over a vector store and creates retrievers from it; its PostgreSQL store is llama-index-vector-stores-postgres.Loadable extensionA shared library that a running SQLite connection loads to gain new functions, types or virtual tables without changing SQLite itself.Local embeddingComputing embeddings with a model you run yourself, so documents and queries never go to an outside API.Local-firstA design in which source files, vectors and the model that embeds queries run on the user's own machine, and anything sent elsewhere is an explicit, visible choice.Logical collectionA table in PostgreSQL or a collection in a vector database: one corpus with its own embedding column, filters and indexes, sharing the engine with others.LookupFetching the row whose key or value matches exactly; there is no ranking because there is nothing to rank.MThe HNSW build setting that fixes how many neighbors each point links to on each layer; more links make the graph easier to search and cost memory and build time.Managed serviceA cloud product where the provider operates the servers, scaling, patching and often the index or pipeline, and you call it through an API or SQL.MarqoA vector search project whose open-source repository is deprecated and no longer receives updates; the company now offers an e-commerce search platform.Matryoshka embeddingAn embedding trained so that keeping only its first m numbers, then renormalizing, still gives a usable shorter vector.Metadata filterA WHERE condition on columns such as tenant, policy_version or status that keeps or drops rows exactly, before the vector ranking runs over what is left.MilvusAn open-source engine for vector search that runs as a Python library, one container, or a distributed Kubernetes deployment.MnemosA TypeScript and Next.js application that answers questions over a person's own files, storing chunks and vectors in one SQLite file with sqlite-vec.MongoDB Atlas Vector SearchVector search over documents stored in MongoDB, on Atlas or self-managed, with hybrid fusion stages.MRRFor each query, 1 divided by the rank of the first relevant row (0 if none), averaged over queries.Multi-queryGenerate several phrasings of one question, retrieve for each, and fuse the ranked lists, usually with reciprocal rank fusion.Multimodal embeddingAn embedding model that maps more than one kind of input, such as images and text, into one space, so a text query can retrieve an image.nDCG@kSum over the top k of (2^grade - 1) / log2(rank + 1), divided by the same sum for the best possible order of the labels.Nearest neighborA stored row whose vector lies closest to the query vector under the chosen distance.NormalizationDividing every coordinate of a vector by the vector's length, so the vector keeps its direction and has length 1.nprobeThe IVF search setting: the number of nearest lists scanned per query; more probes find neighbors in adjacent lists and scan more rows.Object referenceA path or URI to where an original document lives, stored with a content hash and each chunk's offsets, instead of the document itself.OpenSearchAn Apache 2.0 search engine whose k-NN plugin adds vector fields, with Faiss, Lucene and other engines underneath.Operator classThe part of CREATE INDEX that tells PostgreSQL which operator the index can serve; in pgvector it fixes the distance the graph or lists are built with.p95 latencyA latency percentile: 95% of queries finish within this time and one in twenty is slower; it shows the slow tail that a mean hides.Parametric knowledgeKnowledge stored in a model's weights during training: broad, fixed at training time, and without sources.Parent-child retrievalSmall child chunks are searched for precision; the larger parent each belongs to, such as a rule section, is what the model reads.PartitioningSplitting one logical table into pieces by a key such as tenant, each with its own indexes, so a query that filters on the key searches only the matching piece.pgvectorAn open-source PostgreSQL extension that adds vector column types, distance operators and HNSW and IVFFlat indexes, so vector search runs as SQL inside the database.PineconeA vector database offered as a managed service, with namespaces, metadata filters, integrated embedding and full-text search.Policy driftAfter a policy revision, precedents decided under the old version keep looking relevant while the rule behind them no longer applies.Policy fixtureUW-GUIDE, the illustrative written policy the agent's rules and the guidance corpus come from, in versions 2026.2 (current) and 2025.4 (superseded).Policy precedenceCurrent policy outranks superseded policy, which outranks any precedent; untrusted text has no authority.PoolingThe step that turns the encoder's one-vector-per-token output into a single vector for the whole text, for example by averaging (mean pooling) or by keeping a special first token (CLS pooling).Post-filterA filter that removes rows from a result already cut to k, so the result can shrink below its budget, or leak if the step is skipped.Pre-filterA filter that removes disallowed rows before ranking, so the first stage returns its full budget of allowed rows.PrecedentOne decided case with its decision, policy version, date, deciding role, override flag and reason, retrieved to show how similar cases were decided.Precision@kThe number of relevant rows in the top k, divided by k.Product quantizationA compression that splits each vector into sub-vectors and replaces each sub-vector with the id of its nearest entry in a small learned codebook.ProvenanceThe recorded origin of a retrieved row: its source record, who decided or wrote it, in which role, when, and whether it is real or simulated.QdrantAn open-source engine for vector search with payload filters, sparse vectors and a Query API for hybrid and multi-stage queries.QuantizationReplacing each vector by a shorter code (fewer bits per number, or ids into a codebook) so that it takes less memory and compares faster, at the cost of approximate distances.Query planThe database's description of the steps it takes for one statement: which tables it scans, which indexes and sorts it uses, in what order.Query rewritingBefore retrieval, a model rewrites the query into the corpus's wording, or adds terms to it, so the searched text is closer to the stored text.RAG EngineA managed pipeline that ingests files, chunks them, embeds them, indexes them into a corpus and retrieves from it.RAGA pattern in which application code retrieves relevant rows or passages first and a model writes an answer from them, citing what it was handed.Re-embeddingComputing new vectors for existing rows because the vectors would now come out different: a new embedding model or version, or a change to the text that is embedded.Recall@kThe number of labeled relevant rows in the top k, divided by the number of labeled relevant rows.Reciprocal rank fusionA way to merge ranked lists: each list gives a row 1 / (k + rank), rows missing from a list get nothing from it, and rows are ordered by the sum.RecommendationOne of approve, decline, request_info or escalate, with reasons, cited rule ids, evidence ids, the policy version and requires_human set to true.Recursive splitterA splitter that tries separators in order (paragraph, line, sentence, word, character) and splits again only the pieces that are still too long.Redis vector searchRedis's search and query features, which index vector fields next to text, tag, numeric and geo fields and fuse keyword and vector results.Reference architectureA whole retrieval design written down with the decision behind each part, what it gives up, and how to measure that it works, so another team can adapt it instead of copying it.Registered functionA function written in the application's language and registered with the database connection, which SQLite calls once per row when a statement uses its name.ReindexRebuilding an index (an HNSW graph, IVF lists) from vectors that do not change, to repair degradation after churn, apply new build settings or recover recall.Relevance labelA grade (here 2 answers it, 1 related, 0 not) given by a person or by a rule written before any ranking is seen.Rerank top nThe number of candidates kept after reranking (10 by default here); the grounding step then takes the context budget (5) from them.RerankerA scorer that runs after the first retrieval stage, only over its candidate pool, and reorders those rows with signals the first stage did not have; it keeps the top n.Retrieval pipelineA chain of stages (encode, filter, candidates, fuse, rerank, select, generate) in which each stage receives the previous stage's rows and passes on fewer.Retrieval traceA list of entries, one per stage, saying how the stage was configured and how many rows it received and kept, ending with the selected passage ids.RetrievalFinding the rows or passages of a corpus that are relevant to a query, before the model writes anything.RetrieverA framework object that takes a query text and returns a list of documents, with the model, metric, filter, k and query stored inside it.Row idThe stable identifier a vector shares with its source row, so that a match can be turned into a row the agent can read and cite.Score calibrationFitting a function on labeled pairs so that a score means an observed rate of relevance; without it a reranker score orders rows within one query and means nothing more.SelectivityThe fraction of rows that pass a filter; a selective filter keeps few rows, an unselective one keeps most.Semantic cacheA cache that returns a stored answer when a new query's embedding is within a distance threshold of a cached query's embedding, instead of matching the exact text.Semantic chunkingA splitter that embeds sentences, compares neighbors, and places a boundary where the similarity drops, as a sign of a topic change.Semantic searchFinding texts close in meaning to the query, even with few shared words, by comparing vectors.ShardingSplitting data across several machines; a coordinator sends each query to every shard, each returns its own top k, and the coordinator merges them into the global top k.SliceA named part of the evaluation set (a corpus, a policy version, a tenant, a query type) reported separately from the average.Sparse retrievalRetrieval that describes each row by the few vocabulary terms it contains, looks the query's terms up in an inverted index and scores matches, for example with BM25.sqlite-vecA C extension with no dependencies that adds vector functions and the vec0 virtual table to SQLite, loadable from Python, Node.js and other languages.SQLiteAn in-process SQL database library: no separate server, and a complete database with its tables and indexes in a single ordinary file.StemmingNormalizing words by removing suffixes so that forms such as roof and roofs, or differ and differs, become one term.Streaming indexAn index that accepts inserts, updates and deletes while serving, so a new or changed record is searchable without rebuilding the index.SubmissionA JSON record of one broker's quote request, written once by intake from the email; every later step reads it, never the email.Temporal retrievalRetrieval with a date condition inside the query: among the versions of a rule, return the latest one whose effective date is on or before the date asked about.Tenant isolationThe guarantee that a search on behalf of one tenant, here one book of business, never returns, ranks or counts another tenant's rows.TF-IDFA term weight that grows with how often a term appears in a document and with how rare it is across the corpus.TokenA short piece of text, such as a word, a word fragment or a punctuation mark, mapped to an id in a model's vocabulary.TokenizerThe fixed program, shipped with a model, that splits text into tokens and maps them to vocabulary ids.TombstoneA small row that says document X was deleted at time T, kept after the document itself is gone, so every downstream copy can learn about the delete.Top-kThe first k rows after sorting by distance or score; in SQL, ORDER BY distance LIMIT k.Toy encoderA small deterministic function in examples/rag-uw that maps words to 16 named concept dimensions, sums them and normalizes; it is not a model.UnderwriterThe insurer's employee who reads the agent's recommendation and makes the decision.Underwriting agentA program that calls a language model in a loop, runs the tools it asks for, gathers evidence and writes a recommendation for an underwriter.Untrusted contentText such as a broker attachment that retrieval brings into the prompt; it may contain sentences shaped like instructions and is treated as data only.ValdAn Apache 2.0 vector search engine deployed on Kubernetes with Helm; you bring your own vectors.vec0A virtual table that stores a row id and a vector per row and answers KNN queries written as WHERE embedding MATCH ? AND k = ?.VectaraA retrieval service offered as SaaS that generates document and query embeddings itself and blends keyword and vector scores.Vector columnA column of vector type in the table that already holds the data, so each row carries its own embedding.Vector databaseA separate engine that stores ids, vectors and copied metadata and answers nearest-neighbor queries over a network API.Vector SearchAn index service built on ScaNN: you send vectors with ids and filter tokens, deploy the index to an endpoint, and query it.Vector store abstractionA framework's common interface for adding, deleting and searching documents by similarity, whatever engine stores them.vector typepgvector's column type vector(n): n single-precision numbers per row, so every row carries its own vector next to its other columns.VectorA list of numbers that places something as a point in space, so that closeness between points can be computed.VespaAn engine for search and ranking that runs embedders inside it and scores documents with tensors.Virtual tableAn object registered with a SQLite connection that looks like a table to SQL, while its reads and writes run the extension's code instead of the database file's.WeaviateAn open-source engine that combines vector search and BM25 keyword search, with vectorizer integrations that can embed text for you.WHERE clauseThe part of a SQL query that keeps the rows whose columns satisfy a condition and drops the rest, without ranking them.
Each term is explained in full on one page and used everywhere else. Statements about named products and libraries were read in their own documentation on 2026-11-18; the sources are listed on About.