Search With a Book That Does Not Exist

The Library of Congress built a catalog record for a book that does not exist.

The record has a title, subject headings, a classification, and a summary. A language model invents all of them from a researcher's question. In almost any other library workflow, this would be a spectacular failure: confident metadata attached to nothing.

Here, the invention has one narrow job. It is turned into coordinates.

The Library's embedded-catalog experiment begins with about 20.3 million real bibliographic records. Staff extracted descriptive metadata—titles, subject headings, classifications, summaries, tables of contents, and notes—and used Amazon's Titan Text Embeddings V2 model to represent each record as a vector with 1,024 dimensions.

Records that the model treats as semantically similar end up near one another in that numerical space. A search query can be embedded in the same way and compared with the collection.

The researchers tried something stranger. Instead of embedding only the question, they asked a language model to imagine the record that would answer it. The resulting “phantom record” is not added to the catalog or shown as a source. Its fabricated fields provide a fuller bundle of language for locating a neighborhood among the real records. Once the system finds that neighborhood, it returns actual holdings.

The false record is being used as a probe.

That boundary makes the experiment more interesting than a promise of conversational library search. Generated material often becomes dangerous when an intermediate guess crosses into a place that people reasonably treat as evidence: a citation, a customer record, a medical history, a catalog entry. This prototype gives the guess room to be wrong while denying it a shelf.

The same project includes a more conservative tool for catalogers. When a record does not yet have subject headings, the system finds similar existing records and shows the headings people already assigned to them. A cataloger can reuse or modify those terms. The recommendation may still be a poor fit, but it points to real descriptive work rather than generating an authoritative-sounding heading from scratch.

The distinction is small enough to become a practical design test. What is invented? Where is the invention allowed to travel? Which real object must a person reach before the interface asks for trust?

The catalog underneath those questions is not neutral ground. Library of Congress classifications and subject headings carry decades of professional judgment, inherited categories, revisions, omissions, and disagreement. The BIBFRAME initiative is itself a long transition in how bibliographic description is structured and exchanged. Embedding those records does not dissolve that history into pure meaning. It gives a model another way to arrange the descriptions people and institutions already made.

The project's public catalog map keeps both arrangements visible. One view colors the records by Library of Congress class. Another divides them into communities inferred from their positions in the model. The two structures overlap without becoming identical.

Even the picture is an interpretation. The vectors have 1,024 dimensions; the map has two. The Library's pipeline first reduces them to 50 dimensions and then uses a learned projection to place them on a flat surface. Staff explicitly describe that flattening as subjective. A coastline on the map may look inevitable while depending on the embedding model, the selected metadata, the sampled training records, the reduction method, and the labels chosen for each cluster.

A second atlas places the catalog beside English Wikipedia. It can reveal regions where one collection has many records and the other has few, but those gaps do not automatically measure cultural importance or institutional neglect. A catalog record and an encyclopedia article are different kinds of description assembled under different rules. Proximity can make the comparison explorable without making the two collections commensurate.

The model's own documentation adds a useful restraint. Amazon's service card warns that compressed representations can place related concepts too far apart and unrelated concepts too close. Its general retrieval benchmarks cannot establish how well this particular system answers a historian, genealogist, student, or cataloger using this particular collection.

The Library has published an unusually inspectable prototype: the corpus size, model, vector dimensions, database, reduction pipeline, rendering method, public maps, and downloadable embeddings. It has not published a task-level comparison showing when the phantom-record method beats a plain-language query, which false neighbors it introduces, or how researchers understand the results. The post calls the discovery system a proof of concept. That limit should travel with the clever name.

There is a broader use for the pattern, provided the boundary stays sharp. A generated intermediate can help a system explore, translate, plan, or retrieve without becoming part of the final record. The interface can label that material as a working hypothesis, keep it out of durable state, and require the result to land on inspectable sources. The more fluent the intermediate looks, the more important its quarantine becomes.

The next useful test is ordinary. Give researchers the same questions through plain semantic search and through phantom records. Ask catalogers to inspect the neighbors, including the embarrassing ones. Record when the invented description opens a useful path and when it quietly narrows the question to the model's preferred language.

Then keep the real books, the real records, and the failed searches. The phantom record has already done enough by showing where to look.