Skip to content

Glossary

boudou edited this page Nov 8, 2026 · 11 revisions

Glossary

The terms the other pages use, in plain words.

  • Answer contract: what the server tells the client's model to do with an answer: use only the returned passages, cite them, copy facts verbatim, check the draft and show the receipt. See Answers you can prove.
  • Clause path: where a passage sits in its document, such as 7 Termination > 7.2 Notice, taken from its headings and numbered clauses. It is part of each citation.
  • Archived: a citation label, "archived": true, on a passage whose file matches one of its layer's archived patterns (.gitignore style, set with strata_set_archived_patterns or the preset key archived_patterns). See Archived files.
  • Chunk: a passage of a document, cut along its headings or structure. Search returns chunks, each with its file and section.
  • Confidence: high, medium or low, returned with every strata_search result and with a detailed strata_ask response (the concise answer gives status instead); a low one comes with a warning. It is judged on the reranker's relevance probability (0 to 1) or, with reranking off, on the question/passage similarity; the result names the basis.
  • Cut: "cut": true on a clause that came back only in part because the whole did not fit the token budget; strata_ask also says so in its note. See When part of the answer is missing.
  • Document date: the date in a citation, from the document's front matter (date_source: "front_matter") or else the file's modification time ("file"). See Document dates.
  • Embedding: a list of numbers that stands for a chunk's meaning, so similar meanings get similar numbers. The embedding model (all-mpnet-base-v2 by default) makes them; the vector column stores them (EMBEDDING_DIMENSIONS, 768 by default).
  • Extractive answer: an answer made of the document's own sentences, chosen by the server without a language model. See Extractive answers.
  • Grounded answer: an answer your client writes from the cited passages and checks with strata_verify. See Checking a drafted answer.
  • HNSW: the pgvector index that finds similar embeddings quickly; HNSW_EF_SEARCH trades recall for speed.
  • Hybrid search: a search by meaning (embeddings) and a search by words (PostgreSQL full-text search), merged into one list with RRF.
  • interaction_id: the id on each search and ask result while learning is on; pass it to strata_feedback to rate the result.
  • Layer: a named collection of documents, usually one folder or one domain, with its own chunking and routing rules.
  • nothing_checked: the strata_verify result for a draft with citations but no quotes and no facts to check; it gets no receipt, because nothing shows the answer is right.
  • Passage: one cited piece of a strata_ask answer, S1, S2, …, with its exact text and its citation.
  • Query rewriting: optional; a language model turns a question into several searches in the layer's own terms. Learned rewrites are reused without the model.
  • Receipt: a stored record of the quotes behind an answer (rcpt_…), opened with strata_get_receipt, the strata://receipts/{receipt_id} resource or stratakb receipt <id> outside the chat; it shows each quote in its document, marked current, changed (with the newer version) or gone (its file or layer was removed). strata_verify issues them for a verified draft, and extractive answers carry one too. They are kept for RECEIPT_RETENTION_DAYS (90 by default).
  • Replaced by: replaced_by in a citation, naming the documents that say in their front matter (supersedes) that they replace this one. See What a document says it replaces.
  • Reranking: in strata_ask, a second model (a cross-encoder, RERANKING_MODEL) reads the question with each candidate chunk and puts the best first.
  • Routing patterns: a layer's regular expressions. When a question names no layers, those whose patterns match its first 2,000 characters are searched first. A pattern that does not compile or that repeats a repeated part (such as (\w+\s?)*) is refused when the layer is created or updated or the preset is loaded, and skipped if a layer already holds one.
  • RRF (Reciprocal Rank Fusion): how hybrid search merges its two lists by rank, weighted by HYBRID_SEMANTIC_WEIGHT and HYBRID_KEYWORD_WEIGHT, with the constant HYBRID_RRF_K.
  • Token budget: how much passage text strata_ask returns, about 1,500 tokens by default (ANSWER_TOKEN_BUDGET).
  • top_k: how many chunks a search returns: 5 by default for strata_search, 10 per search for strata_ask, whose answer is then cut to the token budget.

Clone this wiki locally