Document Keywords & Summaries

For every PDF and plain-text document in a dataset, the profiler extracts a short list of keywords and a summary of the document's content. Structural metadata — page count, author, title — says what a file is; keywords and a summary say what it is about, which is what dataset discovery searches on.

Both come from a single LLM call per document against the SCAYLE LLM service.

JSON files get keywords and a summary from the same call and the same settings, built from their fields and a few sample records instead of running text; see JSON Profiling.

How It Works

1. The document's text is extracted: PDFs with PyMuPDF (the same pass that produces the CDD chunked_content), text files as read for the text statistics.
2. Documents with fewer than 10 words of extractable text are skipped. Scanned PDFs without a text layer fall here, as do placeholder files; given nothing to describe, the model invents keywords such as placeholder or empty document instead.
3. The text is sent to the model, capped at the first 24,000 characters. Course sheets and teaching material fit whole; longer documents are described from their beginning, where the title, abstract and table of contents live.
4. The model returns up to 8 keywords and a summary of at most 150 words, both in the language of the document. Duplicate keywords (ignoring case) are dropped.

Extraction is best-effort. If the LLM is unreachable, the model id is not served, or the answer cannot be parsed, the failure is logged and the document gets empty keywords and an empty summary — profiling continues. Losing an enrichment must never fail a profile.

Which documents are described

A document gets a record set — and therefore keywords and a summary — only when it sits inside a folder of the dataset, so that the folder is listed as a cr:FileSet:

<dataset>/
└── course_sheets/              ← cr:FileSet
    └── IPB_Offering_example.pdf  ← cr:FileObject + cr:RecordSet (keywords, summary)

A PDF or text file placed directly at the dataset root is listed in the distribution, but has no record set and is not described.

Profile Output

MoMa profile

Keywords and summary appear on the document's cr:RecordSet in the heavy profile:

{
  "@type": "cr:RecordSet",
  "name": "IPB_Offering_example.pdf",
  "source": { "@type": "cr:FileObject", "@id": "55038d8b-2743-42d4-8311-5de74c285bdc" },
  "title": "ECTS Guide Artificial Intelligence",
  "author": "Instituto Politécnico de Bragança, School of Technology and Management",
  "pagesCount": 2,
  "keywords": [
    "Artificial Intelligence",
    "Machine Learning",
    "Python",
    "Scikit-learn",
    "Supervised Learning",
    "Unsupervised Learning",
    "Clustering",
    "Dimensionality Reduction"
  ],
  "summary": "This document describes the Artificial Intelligence course unit for the Bachelor in Informatics Engineering at the School of Technology and Management, academic year 2024/2025. …",
  "field": []
}

In the JSON-LD @context, summary resolves to dg:summary; keywords is schema:keywords.

Not the dataset's keywords

The top-level keywords of the profile are the ones submitted in the profiling request and are never changed. The extracted keywords live on each document's record set.

CDD profile

The same values appear on the document's entry in files, with the summary under description:

{
  "file_object_id": "55038d8b-2743-42d4-8311-5de74c285bdc",
  "original_format": "application/pdf",
  "name": "IPB_Offering_example.pdf",
  "description": "This document describes the Artificial Intelligence course unit for the Bachelor in Informatics Engineering at the School of Technology and Management, academic year 2024/2025. …",
  "keywords": ["Artificial Intelligence", "Machine Learning", "Python", "…"],
  "chunked_content": [ { "section": "", "subsection": "", "chunks": ["…"] } ]
}

Configuration

Extraction is on by default and uses the SCAYLE credentials shared with the other LLM features:

SCAYLE_BASE_URL=https://<scayle-host>/api
SCAYLE_USERNAME=<username>
SCAYLE_PASSWORD=<password>
SCAYLE_VERIFY_SSL=false   # self-signed certificate

ENABLE_DOCUMENT_LLM=true                 # set to false to skip keywords and summaries
DOCUMENT_LLM_MODEL=qwen3.8-flash-next    # default; any id the endpoint serves
DOCUMENT_LLM_TIMEOUT=120                 # seconds per request

Unlike semantic type annotation, this feature does not read llm_config.yaml: that file pins a 30-second SCAYLE timeout, too short for long documents.

Model identifiers are case-sensitive and drift

SCAYLE model ids change over time and are matched exactly. A retired id makes every document come back with empty keywords and summary while the job still succeeds. List the models the endpoint serves before changing DOCUMENT_LLM_MODEL, and check the logs for Document keyword and summary extraction failed, logged once per failed document with the error.

Limitations & Notes

• One LLM call per document, a few seconds each with the default model. A folder of several hundred PDFs adds tens of minutes to the heavy profile; set ENABLE_DOCUMENT_LLM=false when that is not worth it.
• Scanned PDFs have no text layer and are skipped: no OCR is run.
• Long documents are described from their first 24,000 characters only.
• Output is free text, not a controlled vocabulary, and wording varies slightly between runs.