GLOSSARY
Glossary
Thirty-six terms used across LipiCore Learn, one paragraph each, with a link to the article that explains it properly.
access key
A credential issued from your own deployment that an AI client presents to connect to an MCP server. Keys are per team or per tool, can be rotated on a schedule and revoked at any time, and without one there is no connection. The access decision stays with whoever runs the server.
Read: Access keys and revocationaggregation
Any computation over a set of records rather than one: a count per value, a sum, an average, a minimum or maximum. In SQL it is a GROUP BY with an aggregate function; in an index it is a facet, computed on the filtered set and returned beside the results. Most dashboard widgets are aggregations.
Read: What is faceted search?boost
A weight applied at query time that raises or lowers a result's relevance score for a reason the business chooses: a match in the title above one in the description, recent records above old, in-stock items above out-of-stock. Boosting is how a technically correct ranking becomes a useful one.
Read: Full-text search vs database LIKEcollection
A set of indexed records of one kind - products, orders, customers, invoices - with its own fields, its own index and its own refresh schedule. Collections are the unit of organization in an index server: data arrives per collection, and queries name the collection they run against.
Read: What is an index server?doc values
A column-oriented layout inside a search index where each field's values are stored together in record order. It is what makes sorting and faceting cheap: counting how many matching records fall into each brand is one pass over the brand column, with the matching ids already known from the search step.
Read: What is faceted search?facet
A filter that comes with counts. Where a plain filter narrows a result set, a facet also reports how many records each value, range or statistic would leave, computed on the same filtered set as the results and returned in the same request. Four kinds: field, range, stat and nested.
Read: What is faceted search?field facet
A facet that counts records per value of a field: Polo 302, Levis 144, Nike 85 by brand. It is the sidebar count on every catalog, and it is also a GROUP BY: a leaderboard is a field facet sorted by its count, and a pivot is two of them nested.
Read: What is faceted search?freshness
How far an index lags behind its source. An index is a copy refreshed on a schedule, so its contents are as of the last import - minutes to hours, chosen per collection. Freshness is a setting, not a flaw, and the honest way to handle it is to show it: "as of 09:00" on the screen.
Read: Incremental indexing: keeping an index freshfull import
A rebuild of a collection from scratch: every record read once from the source and indexed once. The right move for a first load, after a schema change, or after a large correction in the source. While it runs the previous index keeps serving, so the switch is a cutover rather than a gap.
Read: Incremental indexing: keeping an index freshfull-text search
Search that works on words rather than characters. Text is tokenized, stemmed and stored in an inverted index, so a query is a lookup, results carry a relevance score, typos are matched by edit distance and suggestions come from the data itself. The opposite of a LIKE pattern match, which scans rows and knows nothing about words.
Read: Full-text search vs database LIKEgrounded answer
An answer whose every number came from the data rather than from the model. In indexed retrieval the model maps the question to a request, the index computes the result, and the model writes it up; the numbers are passed through, never produced. The opposite is a plausible answer: fluent, confident and uncheckable.
Read: How can AI answer questions about business data?incremental import
A refresh that brings in only new and changed records since the last run, found by an updated-at timestamp, a change flag or a rising id. It is what runs on the schedule between full imports, and handling deletes - rows that have no timestamp to find them by - is the part to decide before the first run.
Read: Incremental indexing: keeping an index freshindex
A copy of data arranged for a particular kind of lookup. A phone book sorted by name is an index; so is the B-tree a database keeps beside a table, and so is the inverted index a search engine builds over text. Same facts, different organization, different questions answered fast.
Read: What is an index server?index server
A separate system that keeps its own copy of your records, organized for searching, filtering, counting and totalling across millions of them, and serves them through an API. It sits beside the database, which stays the source of truth, is refreshed on a schedule, and is built to be read from, not written to. Also called an indexed data layer.
Read: What is an index server?inverted index
The structure that makes text search cheap: a map from each term to the list of records that contain it, so finding every record that mentions "jacket" is one lookup rather than a scan. A search for two words fetches two lists and intersects them. A database table maps the other way, from a record to its fields.
Read: Database index vs search indexKPI
A key performance indicator: one number that matters, shown with its change against the previous period. On a dashboard it is a tile at the top; in an assistant it is the shape an answer takes when the question asks for a single figure. Orders this period against last is a KPI; the orders themselves are a table.
Read: KPI vs metricLLM
A large language model: the model that reads a question, works out what is being asked, chooses the fields that answer it and writes the result up in plain language. It is good at language and intent and must not do arithmetic or invent values; in a well-built assistant the numbers always come from the index. You choose the provider and the key.
Read: How can AI answer questions about business data?MCP
The Model Context Protocol: an open standard for connecting AI clients to tools and data the same way, so a data source is integrated once instead of once per AI tool. A client discovers a server's tools, calls them with structured arguments and gets structured results back. The plug, not the appliance.
Read: What is MCP?MCP client
The connection an AI application keeps open to one MCP server. The host application - ChatGPT, Claude, Cursor or a company's own assistant - can hold several clients, one per server, and each handles listing the server's tools, sending calls and receiving results. In everyday use "client" and "the AI tool" mean the same thing.
Read: What is MCP?MCP server
A small service that publishes a list of tools - names, descriptions, argument schemas - and answers calls to them with structured results. For business data the right tools search, filter and aggregate over indexed collections; a good server does not publish free SQL or writes. It runs as an endpoint inside your infrastructure, with keys you issue.
Read: MCP server explainednested facet
A facet inside a facet: orders by region, and within each region by category, each level carrying its own counts and totals. Any facet type can nest inside any other, to whatever depth the screen needs. A pivot is a nested facet with two dimensions and a totals column.
Read: What is faceted search?one-way data flow
The rule that data moves from your systems into the index and never back. Applications read from the index; they do not change it, and the index never changes the source. If the two disagree the source is right and the next import fixes the copy, which is why an indexed layer is easy to add and easy to remove.
Read: One-way data flowperiod comparison
Showing a number against the same number for an earlier period - this week against last, this month against the same month last year - inside the same widget rather than in a second chart. On a KPI tile it is the change under the number; on a bar chart, a lighter bar behind each current one.
Read: Choosing the right chart for business datapivot
A table with one dimension down the side, another across the top, and totals: regions by month, with North's 395 orders split across the months and summed in the last column. An assistant builds one from a sentence; an index answers it as a nested facet.
Read: Choosing the right chart for business dataRAG
Retrieval-augmented generation: documents are cut into chunks and embedded, a question retrieves the most similar chunks, and the model answers from them. The right approach when the answer is a passage - a policy, a clause, a procedure - and the wrong one when the answer is a number, because RAG retrieves and does not compute.
Read: MCP vs RAGrange facet
A facet that counts records in buckets of a numeric, currency or date field: price 0-25 (120), 25-50 (245), 50-100 (166); orders this week, last week, earlier. The buckets are computed on the filtered set, so they total the same 531 records as the field facets beside them.
Read: What is faceted search?record
One row of a collection: one product, one order, one customer, one invoice, with its fields. Records are what an index stores, what a search returns, what a table shows, and what facets count. In the database the same thing is a row; in an index it is flattened per collection so that one record answers one line of a screen.
Read: What counts as a record?relevance
A score per result that puts the best matches first, built from how rare a term is across the collection, how often it appears in the record, and which field it appeared in. A database has no equivalent, because a row either matches a predicate or it does not. Boosting adjusts relevance for business reasons.
Read: Full-text search vs database LIKEsaved view
A filtered, sorted dashboard or data grid that a user stores and reopens: the Urgent orders grid at 14, one click from the start of the day. Saved views are how recurring questions stop being questions and become part of the morning, which is the job of a dashboard rather than of chat.
Read: Saved views as a workflowstat facet
A facet that computes a number over the filtered set rather than a count per value: sum, average, minimum, maximum or median. The average price of the 531 matching products, 42.60, is a stat facet, and so is the total value of this month's rejected orders.
Read: What is faceted search?text-to-SQL
An approach where the model is given the database schema and writes SQL from the question. Powerful for an engineer who reads the query, and risky for everyone else: ambiguous joins, queries that run and are silently wrong, every question a load on the production database, and a model that writes SELECT can be talked into writing more.
Read: Three ways AI answers from your data: RAG, text-to-SQL and indexed retrievaltoken
Two meanings, both used here. In search, a token is a unit of text after tokenization - one word, lowercased and stemmed. In language models, a token is the unit of text the model reads and produces, which is also the unit the provider bills by; usage under your own account means those tokens are yours.
Read: Full-text search vs database LIKEtokenization
Splitting text into terms before it goes into an inverted index: words separated, lowercased, punctuation stripped, plurals reduced to a stem so "Jackets" and "jacket" match. Tokenization is why full-text search knows what a word is and a pattern match on characters does not.
Read: Full-text search vs database LIKEtypeahead
Suggestions offered as the user types, matched by prefix on a field built for it: "jac" offers "jacket", "jackets, denim" and "jacquard" before the word is finished. Typeahead comes from the index in the same request path as search, which is what makes it instant.
Read: Full-text search vs database LIKEtypo tolerance
Matching a misspelled term to the one that was meant by edit distance, the number of single-character changes between them. "jaket" is one edit from "jacket" and matches with a fuzziness of one; a fuzziness of two usually adds noise. Spell correction from your own data suggests the corrected term back.
Read: Full-text search vs database LIKEwhite-label
Software that carries your brand rather than the vendor's: your logo, colours and theme on the dashboards, grids and endpoints your team and your customers see. A white-label MCP server, for example, is your endpoint under your name, with keys issued from your own deployment.
Read: White-label analytics