#bfwe2026 Day 2 begins with Matias Frosterus
National Library of Finland on "RDA Aggregates in BIBFRAME: Two Approaches" - methods for fully implementing RDA in BIBFRAME in Finland (BFFI)

*your catalogue will probably never be perfect.

#MetadataNeverSleeps - it's never done because it has to evolve as our understanding of the bibliographic universe & of information evolves.
The problem is resourcing #maintenance to enable evolution when libraries are constantly facing cuts #bfwe2026

An interesting presentation at #bfwe2026 on #Open #Metadata and #BIBFRAME

And that's a wrap on day 1 of #bfwe2026

blue-core-mcp-ingestion skill directs the Agent to create the Work first, then the Instance.
Work goes up first; the agent captures the returned URI, writes it into the Instance’s bf:instanceOf, and creates the Instance second. #bfwe2026

Human cataloguer needs to interpret the Agent's validation report of Work and Instance and give confirmation to continue, modify graphs, or re-run bibframe-transform skill.
Nothing reaches the Blue Core datastore until a cataloguer says yes.
#bfwe2026

The BIBFRAME validation skill uses BIBFRAME Interoperability Group’s demo validation tool.
Validation runs with allow_warnings=True, so only Violations break conformance. These shapes describe interoperability expectations #bfwe2026

Cataloging Judgement is Written Down: FOLIO reference-data UUIDs resolve to labels through the same FolioClient; unlinked controlled values become blank nodes with rdfs:label, not invented URIs.
Agent points at one table row & its triple; the mapping is auditable by a cataloguer #bfwe2026

The mapping from FOLIO to BIBFRAME is produced using 2 markdown tables in the Skills folder.
This enable the Agent to carry FOLIO Instance and Holdings JSON into BIBFRAME Work and Instance JSON-LD, field by field.
There is no transformation script to maintain. #bfwe2026

An example tech services workflow using an AI agent (Claude code) to query the FOLIO Inventory API and retrieve a record then transform it into a Work and Instance in BIBFRAME RDF ready for human approval before addition to the Blue Core MVP #bfwe2026

AI Agents need defined Skills - Agents load only the name and description at startup, then the full instructions when a task matches.
Best practice: the description is the trigger.
Say when to use the skill, not just what it does.
#bfwe2026

Best practice: be specific and correct
exact commands, real file paths, explicit prohibitions ("never commit .env"). A stale instruction is worse than none.
#bfwe2026

AI Agents have harnesses - holding the loop, tools and permissions for the Agent - Commercial buys capability; open source buys control.
#bfwe2026

Agents work in a (ReAct) loop: observe, decide, act, check, repeat - you can describe your goal in plain language and the Agent works out the steps - no coding required.
But humans stay accountable - people approve decisions that matter #bfwe2026

"AI Agents in an Experimental Technical Services Workflow" - Jeremy Nelson, Stanford University
Constructing AI Agents to support tech services workflows - AI Agents act (not just answer); APIs scripts and commands define what they can reach
#bfwe2026

Developing an OCR / Structured Data Extraction extension fro Marva - connects Marva to a phone or web camera to send images to vision supported LLM to extract structured data to populate BIBFRAME records - particularly good for minimising manual keying e.g. in alternate scripts #bfwe2026

Entity reconciliation between Name Authority records and Wikidata - an AI bot can cross-check with basic look-ups to determine a match between entities in different domains #bfwe2026

To achieve the 'network' effect of BIBFRAME, entities and their relationships have to be defined, the quality of the entity descriptions and the defined relationships between them, directly affects how well an LLM-RAG tool works #dataquality #bfwe2026

Bib records are well suited to RAG methodology; compact and rich data - an LLM-RAG tool in discovery can be built with guardrails so it doesn't hallucinate resources that do not exist #bfwe2026

Interesting visualisations of the metadata for LoC Collections, such as a timelapse of classification over time #bfwe2026