Universidad ICESI: automatic metadata extraction with MetaExtract API on its own infrastructure

Institutional context

Universidad ICESI is a private university in Cali, Colombia, with a recognized track record in research and a library that supports both teaching and the academic output of its centers and research groups. Its information management team already operated its own consolidated infrastructure, with systems chosen and in operation.

The challenge

The problem was not the platform but capacity. Document description is linear work: every incoming document requires a person to read it, identify its descriptive elements and transcribe them field by field. When the volume of documents grows faster than the team, the result is predictable: a queue of material awaiting description that never shrinks, and a high opportunity cost, because the people doing the transcription are trained library professionals who could be working on authority control, content analysis or support for researchers.

The university's condition was clear from the start: no change of system, no content migration and no disruption of existing workflows. What it was looking for was a component that would slot into the current workflow and take transcription out of the way.

What we implemented

We implemented MetaExtract API, our service for automatic metadata extraction from PDF documents using artificial intelligence. The API receives the document, analyzes it, identifies its descriptive elements (title, authorship, abstract, keywords, date, document type, language and any additional fields the institution defines) and returns them in structured form to auto-fill the description form.

Because it is a service delivered through an application programming interface, it required no new infrastructure and no replacement of systems: it integrates with what the institution already has.

How we did it

1. Assessment of the target metadata schema. We reviewed the fields in use, which ones are mandatory, which vocabularies apply and which document types needed to be distinguished.
2. Tuning the extraction. The model was trained with real documents from the institution, so that its output would match the institution's schema and not a generic one.
3. Integration. The API was connected to the workflow, coordinating credentials, usage quotas and error handling with the university's technical team.
4. Validation with a control batch. A representative set of documents was processed and compared field by field with the description produced by the team before general use was opened.
5. Production launch and follow-up. The human review procedure prior to the approval of each record was defined, and the service was opened to the team.

Result

Description work has moved from transcribing to reviewing: the librarian receives a proposed record and applies professional judgment to name normalization, descriptors and licensing decisions, which is where their training adds value. The most significant indicator of the result, and the only one we can state without the university publishing figures, is that the service has been renewed for consecutive terms: the institution evaluated its continuity and decided to keep it.

Technologies

MetaExtract API, metadata extraction with language models, integration through a REST API, Dublin Core metadata schemas and the institution's local fields.