Lesson 5
Inspect a corpus
A corpus is an explicit set of local members. It is not a directory scan and it does not follow URLs, read standard input, discover files, or apply changes. The specification states which sources and verified approved revisions belong to the inspection.
What you will learn
You will learn how corpus inspection changes the question from one document to patterns across documents, why membership must be explicit, and how to read a complete or partial result without treating it as a hidden quality grade.
Move from one document to patterns
Document preparation asks, “What happened to this document?” Corpus inspection asks, “What can we observe across this declared set of members?” The answer is an aggregate report, not a replacement for individual document records.
Corpus inspection aggregates source-text-free evidence. Its static report can group member status, extraction deltas, findings, and listed revision histories by family, format, rule, or severity. It helps you ask pattern questions:
- Which document formats show a condition?
- Which extractors differ most often?
- Which approved revisions are included?
- Is the inspection complete, partial, or failed?
The report does not copy decision authority from a refinement. It does not embed arbitrary source passages or recommend a change.
Why explicit membership matters
An explicit corpus specification names each member and its local source. This makes the collection inspectable too. A reader can see what was included and what was not. A later file that appears in the same directory cannot silently become part of the result.
The same principle applies to revisions. A corpus can list an approved revision only when its associated refinement, diagnosis, and base evidence verify together. A rejected refinement is useful evidence about a decision, but it is not a prepared revision and cannot be included as one.
Understand complete and partial work
The committed corpus examples contain PDF members. A complete run needs local Docling models. If those models are missing, usable non-PDF members can still produce a valid PARTIAL corpus record. This means that the report honestly describes an incomplete declared collection.
An invalid listed revision is different. It stops publication instead of becoming a partial result, because the requested evidence relationship cannot be trusted. Learn to distinguish these two cases:
| Result | Meaning |
|---|---|
COMPLETE | All declared members reached a usable result. |
PARTIAL | The report records both usable members and member-level failures. |
| Failed publication | The requested inputs or evidence relationships were invalid. |
None of these statuses says whether the corpus is suitable for every downstream task. They describe what this inspection could establish.
Try it with the CLI and Workbench
After downloading the local Docling models, inspect the quality corpus:
docling-tools models download layout tableformer \
--output-dir .cache/docling/models
corpus inspect fixtures/corpus/quality-corpus.json \
--docling-artifacts .cache/docling/modelsKeep the Workbench open and select Refresh workspace. Refresh discovers records that the CLI published after the Workbench started. It does not run a new observation, diagnosis, or corpus inspection.
Choose the published corpus in the Corpora list. Read its totals and member evidence in the central area. Then use the Inspector to view its summary, evidence, and artifacts. The corpus view does not show a document stage navigator because it answers questions about several members, not one document lifecycle.
What to remember
Corpus inspection makes a declared collection inspectable. It summarizes the evidence that already exists; it does not discover a collection, change a member, or replace the evidence for an individual document.