Glossary

Taking a pre-trained model and continuing to train it on a narrower dataset so its behaviour shifts towards that domain.

Any model whose output is newly produced content — text, image, audio — rather than a classification, ranking, or extracted span.

A generative model's output that is fluent, confident, and wrong — a plausible-sounding citation, date, or quotation that does not exist in any source.

A very large neural network trained to predict the next token of text — the thing meant by 'AI' in most 2020s headlines.

The text you hand to a generative model to shape its output — instructions, context, worked examples, and the user's question.

A pattern where a generative model is first fed the results of a search over your documents, then asked to write an answer grounded in what it retrieved.

Training is how a model learns — done once, expensively. Inference is how the model is used — each prediction is a forward pass over frozen weights.

A model's own estimate of how sure it is about one output — one token, one region, one classification — used to decide what a human should double-check.

A list of numbers — usually 384 to 1024 dimensions — that represents the meaning of a chunk of text, placing similar passages close together in that space.

The specific neural network that turns a chunk of text into its embedding vector. Different embedding models produce different geometries from the same text.

The step before OCR that identifies and labels regions on a page image — text blocks, headers, footnotes, figures, tables, marginalia.

A trained function from input to output. Umbrella term — an LLM, an embedding model, an OCR engine, and a classifier are all models.

An embedding model trained so that text meaning the same thing lands nearby regardless of language, letting a query in one language retrieve passages in another.

Embeddings produced by table lookup rather than a transformer forward pass. Orders of magnitude faster, with no sequence-length limit.

Splitting text into the sub-word pieces a model actually sees. Token length — not word count — is what drives cost and context-window limits.

The neural-network architecture, introduced in 2017, that sits behind nearly every modern language, vision, and multimodal model.

A model that takes an image (or image + text) as input and produces text — describing, transcribing, or classifying what it sees.

A classifier that assigns a likely natural language (or script) to a span of text, so tools like embedding models aren't applied to text they weren't trained on.

A pair of techniques that find near-duplicate texts in a corpus without comparing every pair — a compact fingerprint plus an index for locating near matches.

Tagging spans of text that refer to people, places, organisations, dates, and similar categories — the step behind entity facets and co-occurrence analysis.

Splitting a stream of prose into sentences. The rules look simple until you meet historical punctuation or OCR noise.

Using measurable features of writing — function-word frequencies, sentence lengths, punctuation — to profile authorship.

A classical score for a word's importance to a document — high when the word is frequent *here* but rare across the corpus.

An unsupervised technique that groups documents or passages by the themes they share, surfaced as lists of related words per topic.

Writing a text's sounds in another script rather than its meaning. Not translation — a reader learns how a word sounds, not what it means.

The standard ranking function for lexical search — what Lucene, Tantivy, and Elasticsearch use under the hood.

Grouping items by similarity without pre-specifying what the groups should be. An unsupervised view of the corpus.

A number between -1 and 1 that measures the angle between two vectors. 1 is identical direction, 0 is unrelated.

Question answering that returns a span from a real document. It can pick the wrong span — it cannot make one up.

A clustering algorithm that finds clusters of varying sizes in dense regions and leaves outliers unlabelled rather than forcing them in.

Running lexical and semantic search over the same query at once and fusing the two ranked result lists, rather than picking one engine.

Reading characters off a page image and producing machine-readable text, line by line, with a confidence score per token that downstream analysis can filter on.

A method for combining two ranked result lists — lexical and semantic search, say — into one ranking, using only each result's rank, not its raw score.

Search that ranks by meaning rather than literal word overlap — useful when the corpus's vocabulary isn't your query's vocabulary.

A technique that squashes high-dimensional vectors down to 2 or 3 dimensions for visualisation, trying to preserve neighbourhood structure from the original space.

Finding the k chunks whose embeddings are closest to a query embedding. The mechanical step behind semantic search and neighbour lookup.

Collective benefit, Authority to control, Responsibility, Ethics — principles for Indigenous data governance that complement FAIR by asking whom data serves and who decides.

Finding and structuring citations — footnote markers, bibliographic entries, in-line references — so they become queryable objects.

Reading one text carefully, in full, attending to its actual language. The traditional method, and still the one that produces most arguments.

Two entities (or terms) appearing together within some window — sentence, paragraph, document — are co-occurring.

The movement to make library, archive and museum collections usable for computational research — as text, data and metadata — not only as items to view one at a time.

A bounded collection of texts gathered to answer a research question — not just 'all my PDFs'. It becomes one once scope and provenance are known.

Analysis that tracks change across time — vocabulary, concepts, named-entity mentions, sentence structure — by slicing a corpus into buckets and measuring each.

Analysing a corpus too large to read, by computing over it. Franco Moretti's term, and a deliberate provocation.

Four principles for research data — Findable, Accessible, Interoperable, Reusable — now widely expected in research funding and data management plans.

Regions — much of Africa, Asia, Latin America and Oceania — shaped by colonial histories and unequal access to resources; in digital humanities, a lens on whose archives are digitised, and for whom.

The record of where something came from. In DH, the chain of custody of a source; in computing, which model produced which output.

A specific convention for writing one script in Latin letters. IAST, ITRANS and Harvard-Kyoto all romanise Sanskrit, and all disagree.

Archēglyph's contract for adding a new analysis technique. A single Python file that declares its dependencies, cost tier, and the UI surfaces it extends.

Our internal library that owns every bundle read and write. Every pipeline step that changes a dataset goes through it.

The per-dataset artefact that holds everything Archēglyph needs to serve a dataset — Tantivy index, zvec index, and metadata SQLite.

The unit of analysis inside Archēglyph — a passage of several sentences, small enough for an embedding model to handle in one pass.

The short title and one-sentence summary a language model generates to describe a cluster's fragments — a description of a grouping, not a claim about the archive.

A plain-language summary of what a dataset contains and how its current results were produced — auto-generated, owner-editable.

A dataset whose owner has turned on public visibility. Gets a canonical URL and a license; some surfaces open up, others stay login-gated.

Named, immutable snapshots of a dataset's bundle at a point in time. Let you return to exactly that state later.