Loading page…
The experiment retrieves existing subject headings and uses fictional records as search aids. Its visualizations show relationships through one model’s lens—not an objective map of knowledge.
Listen to this story
The Library of Congress is testing whether semantic embeddings can make its catalog easier to search and explore, not indexing the books themselves or launching a public service. The sandbox encodes 20.3 million metadata records through the end of 2025 as 1,024-number vectors using Amazon Titan Text Embeddings V2 on AWS Bedrock. In one discovery workflow, an LLM invents a catalog-style “phantom record” from a research question, then uses it to retrieve real records; the resulting visual clusters are model-dependent, not objective classifications.
A subject assistant suggests headings by retrieving similar records and surfacing headings catalogers can reuse or modify; staff are testing it in a BIBFRAME editor.
The phantom record is fictional and serves only as a search aid; the matches returned to researchers are real catalog records.
The interactive map can display records by Library of Congress Classification or by model-generated clusters, but its two-dimensional layout can create arbitrary-looking groupings.
A research question can lead to real library resources through a catalog record that never existed. That is one discovery method the Library of Congress described on October 26 in an AI Sandbox experiment. Staff converted approximately 20.3 million catalog records into numerical representations to explore search, cataloging assistance and visual maps of the collection.
The experiment covered bibliographic records through the end of 2025, not the full contents of the works they describe. Staff extracted descriptive metadata: titles, subject headings, classifications, summaries, tables of contents and notes. Amazon Titan Text Embeddings V2, accessed through AWS Bedrock, converted each record into a vector—a set of 1,024 numbers representing the text.
Those vectors act like coordinates, placing related descriptions near one another. Different terms for the same thing, such as cars and automobiles, can be close together; the model can also connect related text across languages. Staff stored the vectors in an AWS S3 Vector bucket, where new text queries could retrieve similar catalog records.
One application starts with an unfinished catalog record that has no subject headings. A subject assistant converts it into a vector and retrieves similar records. Catalogers can then inspect the headings attached to those existing entries and decide whether to reuse or modify them. The recommendations draw on actual catalog records rather than newly generated headings.
Staff found the assistant useful for surveying which headings were already being used in a subject area, according to the Library’s account. They have been testing recommendations inside their BIBFRAME cataloging editor, allowing catalogers to see suggestions while working. This is a cataloging experiment, not a claim that the model should assign headings on its own.
The discovery proof of concept takes a different approach. A large language model turns a researcher’s question into a fictional “phantom record,” complete with a title, subject headings, classification and summary. Staff then embed that record and search for its nearest matches. The invented record is a search aid, not a resource presented to the researcher as real.
The distinction is deliberate: the fictional details describe where a useful resource might sit among the catalog’s vectors. The returned matches are real records. The Library describes this as an area for continued exploration in discovery systems using BIBFRAME, rather than a completed public search service.
The same vectors also became an interactive map of the catalog. One view colors records by their human-assigned Library of Congress Classification. Another groups them into communities based on how they cluster in the model’s representation. Users can zoom in and inspect records, comparing the model’s organization with the Library’s classification system.
Flattening 1,024 dimensions into two creates a subjective view, the Library cautions, and can produce seemingly arbitrary groupings. The map therefore reflects this particular embedding model, not an objective arrangement of knowledge. Its time-lapse views add another perspective: users can follow records by catalog-entry date or by the publication dates of the resources.
For a comparison beyond the catalog, staff embedded seven million English Wikipedia articles using the same model. The resulting overlay shows both overlap and differences. Wikipedia’s many articles about individual sports players, for example, occupy territory without a matching collection of thousands of individual player biographies in the Library’s catalog.
That comparison is not one-to-one: an encyclopedia article and a catalog record hold fundamentally different kinds of information. The Library suggests comparing other catalogs as another application. For researchers who want to examine the approach themselves, its post also provides embedding download links and technical details about how the visualizations were built.
Loading discussion...
Join the conversation
Explain whether returning real sources makes that approach reassuring—or still uncomfortable.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.