Skip to content
Acervo
Help articles
  1. 1. Getting Started
  2. 2. Workspaces & Teams
  3. 3. Uploading Documents
  4. 4. Processing & Outputs
  5. 5. Reviewing in the Workbench
  6. 6. Curating the Batch Roll-up
  7. 7. Delivering & the ArchivesSpace Integration
  8. 8. Exporting Your Results
  9. 9. Credits & Billing
  10. 10. Privacy & Data Handling
  11. 11. Trying Acervo without an account

Processing workflow

Processing & Outputs

See what happens during processing and what outputs Acervo generates.

On this page

Section links appear after the article loads.

Processing & Outputs

When you click Start processing, Acervo runs a series of activities on your documents. They happen in a deliberate order, with the output from one feeding into the next, but you don't need to choreograph any of it. This section describes what's happening behind the scenes between upload and the workbench.

You can leave the page; processing continues server-side. Single-document jobs typically finish in less than a minute. Processing time increases with batch size and depends on page complexity and whether you've selected Standard or Premium.

Reading the document

Acervo extracts text from your images using a Gemini vision model. The output isn't just a flat string of letters. It's structured text that preserves reading order, paragraph breaks, and page boundaries. Every later activity in the pipeline draws from this OCR output.

For multi-column layouts that belong to one intellectual unit (a printed report, a pamphlet), the model preserves the reading order across columns. For pages that are actually multiple unrelated articles (like a newspaper page), the model will conflate them. Process those clipped articles individually instead.

Identifying named entities

From the OCR text, Acervo identifies the named entities that appear in the document

  • Person · individual people named in the document
  • Organization · institutions, companies, government bodies, associations
  • Venue · named performance and event spaces (theaters, halls, civic buildings)
  • Places · physical places, specifically named locations
  • Work · titled works (books, songs, oratorios, exhibitions, etc.)
  • Event · named events (concerts, conferences, ceremonies)
  • Reference · referenced documents, accession numbers, archival identifiers

Each entity is captured with the form it takes in the text, a normalized form (e.g., "P. Casals" normalized to "Pablo Casals"), and the page where it was found. Entities that span multiple pages are kept linked across pages so Acervo can deduplicate them later.

Generating titles and scope and content

With OCR text and identified entities in hand, Acervo generates the narrative description of the item. A title (DACS 2.3 / EAD unittitle) and a scope and content note (DACS 3.1 / EAD scopecontent).

Where the source document has a clear formal title (like a printed report's title page), Acervo transcribes it. When there's no formal title, Acervo constructs a devised title from the document's content and named entities.

Matching entities to authority files

Each entity is matched against external authority files where applicable

  • Library of Congress (LCNAF for names, LCSH for subjects)
  • VIAF · the Virtual International Authority File, which clusters identifiers across national authority files
  • GeoNames · for places
  • Wikidata · broad-coverage knowledge base

Matching uses three tiers, ordered cheapest to most thorough

  1. Prescreener · fast heuristic checks (type match, temporal/geographic compatibility, description-context conflict, generic-concept rejection). If all checks pass, the match is accepted directly.
  2. Semantic validator · when the prescreener flags concerns, a lightweight language model holistically validates the candidate, weighing biographical and contextual details.
  3. Alternative discovery agent · if the validator excludes the candidate, an agent searches Wikidata for better alternatives, evaluates them with an LLM, and accepts the best one above a confidence threshold.

If all three tiers fail to find a confident match, Acervo does not commit an authority match for that entity. The entity stays in the record without a link. When a match is committed, the entity links out to the authority file directly so you can sanity-check it with one click.

This conservative approach is intentional. False matches in authority records cause real downstream problems for catalogs and discovery layers. Acervo would rather leave an entity unmatched than confidently merge two distinct people.

Applying subject headings

Acervo applies topical subject headings from controlled vocabularies

  • FAST/OCLC · Faceted Application of Subject Terminology, OCLC's faceted derivative of LCSH
  • LCSH · Library of Congress Subject Headings, the foundational subject vocabulary for archives and libraries

Each subject heading carries a confidence score and a rationale explaining why the model picked it. Hover the heading in the workbench to see both. Next to the heading, FAST and LCSH chips show the source vocabulary's identifier on hover (e.g., FAST 1030852) and click through to the authoritative entry, so you can verify the heading directly against the controlling vocabulary.

Scoring confidence

Acervo surfaces confidence scores in two places

  • On named entities, built from two questions. How clearly the source document supports the entity (Source Text) and how legible the underlying image is (Visual Legibility).
  • On subject headings (FAST/OCLC and LCSH), each accompanied by a rationale you can read on hover.

Scores are not surfaced for authority matches. Authority is binary in the workbench. A match is either committed (with a hyperlink to the authority file so you can verify it directly) or not committed. There is no aggregate confidence score at the record level either; scores live at the entity and subject-heading level where the cataloger can act on them.

See §5 Reviewing in the Workbench for how to triage your output using these scores.

Aggregation across pages

For multi-page jobs, Acervo does a final pass to consolidate everything

  • Entity deduplication · recognizes that "Pablo Casals" on page 1 and "P. Casals" on page 4 are the same person, and merges them into a single record
  • Scope and content consolidation · combines per-page descriptions into one coherent note covering the whole document

This is why the page order you set during preview matters. Acervo uses cross-page context (entities mentioned earlier, references back to earlier pages) when it aggregates.

Where the outputs surface

Once processing completes, you'll find everything in the workbench

  • Titles and scope and content
  • Named entities, each with its confidence score and (where matched) a hyperlink out to its authority file for verification
  • Subject headings, each with a confidence score and rationale on hover, plus FAST and LCSH chips that show the identifier on hover and link out to the authoritative entry
  • Source OCR text, available for export

The workbench is where judgment happens. Review the output, edit titles and descriptions, fix wrong authority identifiers at the source, and curate what delivers. See §5 Reviewing in the Workbench and §6 Curating the Batch Roll-up for the workflow.

Reprocessing

Reprocessing is always user-initiated. Acervo never reruns a job on its own. If you want to apply a newer model, an updated vocabulary, or just take another pass, you'll start a fresh job by re-uploading the source documents. Acervo deletes the originals after each processing run, so the source has to come from your archive each time you reprocess.

Past jobs and their outputs stay accessible in your job history for reference and export. See §8 Exporting Your Results for the reprocessing flow, and §10 Privacy & Data Handling for what's retained vs. what's deleted.

Where to go next