How Lyra’s AI features handle your library’s data

Last updated: September 8, 2026

Lyra uses two AI services to power features in the staff app and the patron-facing OPAC:

• Anthropic’s Claude API powers Vega, our conversational analyst in the staff app. Claude generates Vega’s responses to librarian questions.
• OpenAI’s embeddings API powers semantic search across your catalog, both in the staff app and in the OPAC. OpenAI generates a numerical representation (“embedding”) of each catalog item, which Lyra uses to find similar items by meaning rather than exact word match.

This page is for school administrators, data privacy officers, and IT teams who need to understand what data is and isn’t sent to these services when your library uses Lyra.

The short version

AI-powered features are part of Lyra’s current service design and are not separately disabled at the account level. The safeguards described on this page apply to those features by default.

Lyra sends Anthropic and OpenAI information about your collection — what’s in your catalog, how it’s organized, how often items circulate. Lyra does not read your patron records into either service: no student or teacher names, IDs, emails, grade levels, or individual borrowing history is pulled from your database into an AI request.

That part is enforced in our code, not just in policy. The data structures that carry your library’s records into both APIs are designed so that patron-identifying fields cannot enter them.

There is one path our code does not control: free text that a person types. A librarian’s question to Vega and a patron’s OPAC search box are sent as written. If someone types a student’s name into either one, that name is sent — we do not filter it out, and for Vega it is also stored with the conversation and reused as the conversation’s title. Vega is instructed to decline patron-specific questions, but that instruction takes effect after the question has already been sent. Tell your staff and patrons not to type patron names into Vega or the catalog search box.

What we send to Anthropic

When a librarian asks Vega a question, Lyra may send the following to Anthropic’s Claude API:

• The text of the librarian’s question, exactly as typed (see the caveat above — Lyra does not strip names from it)
• The library’s name and basic configuration (grade range served, library defaults)
• Catalog metadata: item titles, authors, ISBNs, publication years, tags, target audiences, reading levels, copy counts, call numbers
• Aggregated circulation statistics: total checkout counts per item, last-circulated dates, current hold counts, turnover rates by collection segment, dead-inventory counts
• For acquisition suggestions, search queries to Anthropic’s web search tool, scoped to bibliographic topics (e.g., “2024 Newbery Medal honorees,” “middle-grade graphic novels about friendship”)

What we send to OpenAI

Lyra sends data to OpenAI’s embeddings API in two situations.

When catalog items are added or updated. Lyra sends the item’s title, authors, description, and topics. OpenAI returns a numerical embedding, which Lyra stores. The original text is not retained by OpenAI beyond the brief period needed to generate the response.

When a search is performed using semantic matching. This happens in two places:

• When a patron searches the OPAC, Lyra sends the patron’s typed search query to OpenAI to embed it for matching against the catalog.
• When Vega uses its semantic search tool to answer a question, Lyra sends the search term Vega generated to OpenAI.

In both cases, the search text is sent on its own, without any identifying information about the patron or librarian making the search. OpenAI has no way to attribute a search query to a specific user. Lyra does not intentionally include patron-identifying information with semantic-search queries. Search queries are sent without account identifiers, patron identifiers, IP addresses, school-user identifiers, or borrowing-history context. Users should not enter personal information into catalog search fields.

OpenAI is used only for items and search queries. Patron records, circulation records, and library configuration are never sent to OpenAI.

What we never take from your patron records

Lyra never reads the following out of your database into a request to Anthropic or OpenAI. This list describes what our code does with your records; it does not describe free text a person types into Vega or a search box, which is covered above and is sent as written.

• Patron names, preferred names, or full names
• Patron IDs, barcodes, or email addresses
• Notification email addresses for parents or guardians
• Homeroom or class names
• Individual borrowing history tied to a specific patron
• Class year values (Vega uses grade-level ranges, not specific class years)
• Any free-text notes or fields that could contain patron-identifying information
• Account identifiers, session identifiers, IP addresses, or other technical identifiers that would identify the patron or staff user who initiated a search or Vega request.

Patron OPAC search queries are sent to OpenAI to power semantic search, as described above. They are sent without any identifying information about who typed them, but the query text itself is sent as written.

If a librarian asks Vega a patron-specific question, Vega is instructed to decline and point them to the relevant view in the staff app, which doesn’t involve AI. That is a behavioral instruction to the model, not a technical guarantee, and it happens after the question has reached Anthropic.

What our AI providers do with the data

Anthropic and OpenAI are Lyra’s subprocessors for the AI features described above.

Anthropic. Under our commercial agreement, Anthropic does not use our API data to train its models. API data is retained for 30 days before automatic deletion. Anthropic is contractually bound to use the data only to provide the API service Lyra has contracted for. Anthropic’s current data handling commitments are documented at https://privacy.claude.com/en/articles/7996866-how-long-do-you-store-my-organization-s-data.

OpenAI. Under OpenAI’s standard API terms, API data is not used to train OpenAI’s models by default. Lyra has confirmed this setting on our account and has executed OpenAI’s data processing addendum. API data is retained for a limited period before deletion. OpenAI’s current data handling commitments are documented at https://openai.com/enterprise-privacy/.

Web search

When Vega is helping with acquisition recommendations, it may run web searches to find currently popular or award-winning K–12 titles. These searches are bibliographic in nature — titles, authors, genres, award lists. Web-search inputs are structured fields rather than free text, and a server-side sanitizer rejects any input containing your school’s name, a homeroom name, or a name from your patron records. This sanitizer runs on web-search inputs only — it is not applied to the librarian’s question itself, or to semantic catalog search.

Audit and oversight

Vega conversations are stored in your Lyra database while they’re active. Lyra retains your most recent 250 Vega conversations per library. Older conversations are automatically deleted. Librarians can delete their own conversations, and conversations are removed when the associated user account or library is deleted. If your school requires a long-term audit record of Vega activity, contact us — we can arrange retention separately.

FERPA posture

Lyra holds student education records on behalf of our school customers and operates as a “school official” under FERPA’s school official exception. Our contracts with schools bind us to use education records only for the purposes the school has authorized, and to flow those same restrictions through to any subprocessor — including Anthropic and OpenAI.

Lyra’s AI features are designed to exclude patron-identifying information from everything Lyra reads out of your education records and sends to our AI providers, so that data does not carry personally identifiable information. The exception is free text a user types, which is sent as written and can contain whatever they put in it. The subprocessor relationships are disclosed and both providers are bound by our standard subprocessor terms; the practical exposure is collection metadata, aggregate circulation statistics, topical search queries, and whatever your staff and patrons type into a search box or into Vega.

If your school requires additional documentation for a vendor review, our data processing agreement and subprocessor list are published on this site.

Questions

If you’re evaluating Lyra for your school and have questions this page doesn’t answer, contact support@lyralibrary.com.