Recently the staff of the Network Development and MARC Standards Office at the Library of Congress participated in an AI Sandbox program hosted by the Library’s Digital Strategy Directorate in the Office of the Chief Information Officer. The Sandbox provided a space to use AI and machine learning tools and methodologies with the Library’s collections and metadata at scale.  

One area we were particularly interested in is vector embeddings and how they can be leveraged with the Library’s catalog records. In this blog post, we will review what embeddings are, go over some of the areas where we experimented with them, and explore some visualizations that allow us to see the approximately 20 million records in the Library’s catalog and how they have developed through time. 

Vector embeddings, in the simplest terms, are numbers that represent text. They are used in many machine learning and AI tools and provide a way to reduce natural language to numbers, which enables us to interact with language using algorithms. This blog post on the collaborative platform and community hub Hugging Face provides a detailed history of vector embeddings.  

In our case, we wanted to generate vector embeddings for each of the 20 million records in our catalog. To do this, we used an embedding model, a neural network–based model like a Large Language Model (LLM), but instead of generating text, it generates vector embeddings.  

Each of these vectors represents a location in the model, and they act as a sort of coordinate system, so two vectors that are in close proximity indicates that they are related. The idea of “closeness” is a bit different from what we would naturally think of in this case. Saying two things are close in physical space means they are in proximity in three-dimensional space. But these vectors represent what is called high-dimensional space (in our case 1,024 dimensions) with dimensions representing abstract variables or features of the text

One useful aspect of having this high-dimensional space is that language is just one feature of the text. This means the model is multi-lingual, and text about the same things in two different languages will still be in proximity to each other. The same concept applies to the terminology and vocabulary used. For example, the word “cars” and the word “automobiles” would be close together — even though they are different words, they are still about the same thing. 

To generate our vectors, we used Amazon Web Services’ embedding model called Titan Text V2. Each catalog record was prepared by extracting only the descriptive metadata — things like subject headings, Library of Congress Classification, summary and table of contents, notes, etc. — and run through the model to generate a vector for each record. These vectors were then loaded into a vector database, which allowed us to query using new text and find the records that are most similar. 

Having the catalog in this format allowed us to develop many different types of potentially useful applications. The first one we experimented with is a subject assistant. The idea is fairly straightforward: take a record that is being cataloged but does not yet have subject headings, turn it into a vector, and then see what is most similar in the database to this in-progress record.  

Once you have a few records that are similar to it, you can show what subject headings were used on the  related records and potentially reuse or modify existing subject headings already applied. In our testing, we found this can be a useful tool for cataloging, helping speed up the process of evaluating what subject headings are currently in use for a specific domain, or getting the lay of the land for subject usage.  

This embedding-based approach only uses real records, so unlike generative AI, there is no risk in generating false or hallucinated subject headings. We have been experimenting with this tool by integrating it into our BIBFRAME cataloging editor, testing its ability for catalogers to see potential recommendations for subject headings while they are working. 

Another interesting application is discovery. We experimented with a proof of concept where a research question was transformed into a “phantom record” — one totally invented by a large language model but perfectly encapsulating the researcher’s question. The LLM generates a title, subject headings, LCC (Library of Congress Classification) headings, summary, and more. It doesn’t matter that it invents this information because the goal is not to use it. The goal is to find where in the embedding model and our existing records it would be located.  

This phantom record is then turned into a vector and queried against the database, and the results are real records that best match the researcher’s question. This is an area that could potentially be very useful for discovery systems using BIBFRAME, and we hope to continue to explore it and see its possibilities. 

Another advantage of having the catalog in this vector form is that it lends itself to data visualization. Because the vectors represent, in a sense, the coordinates of where a concept is located within the embedding model, you can organize the records into discrete regions and represent that visually. However, visualizing vector embeddings is challenging because you are reducing, in our case, 1,024 dimensions into two dimensions (X and Y) so you can plot them visually. This flattening produces a subjective view of the data since we cannot visualize 1,024 dimensions.  

The Library’s rich catalog data also carries human-assigned classifications, such as LCC. This gives us not only an organizational system for where a concept sits in the model, but also for how the Library has cataloged the resource, whether it’s a book about biology, a folk audio recording, or a novel. These human-assigned classifications can be used in the visualization as well.  

Combining these two ideas — the model’s internal structure and the applied Library of Congress Classification — can build a visualization that reflects not only how the model represents these records, but how we view them as well. From there, we can see how our classification systems naturally align with the model’s internal organization.  

The visualization below shows all 20 million records, each placed at its location within the model and color-coded by its LCC class. The results are highly structured groupings of similar resources. The visualization is really an interpretation of the catalog through the lens of the embedding model, but it has the added benefit of the LCC classification, providing labels and color structure as well. 

LINK TO INTERACTIVE VIZ 

We can use the mouse to pan around and zoom in, all the way down to the individual record level. We can also click to see what types of records are situated in that area of the visualization. This is just one view of the catalog — a view through the particular embedding model we used — but it is powerful enough to be able to see the entire 20 million records at once, along with where they aggregate and form structures. 

The visualization also has a “Communities” toggle that provides a view of the catalog in line with the embedding model’s internal organization. This mode divides the resources into communities based on how they cluster together using an algorithm called Hierarchical Density-Based Spatial Clustering of Applications with Noise (HDBSCAN). A process then names each cluster and sub-region in the community based on sampling the titles and subject headings in that cluster.  

Instead of overlaying the LCC classification and trying to force the visualization into that system, this mode reflects the model. It looks more meticulous than the nebulous LCC version, but the reduction from 1,024 dimensions to two can produce groupings that are seemingly arbitrary. 

In addition to this overall view of the catalog, we can also introduce a temporal element to the visualization by limiting which records are shown based on when they were entered into the catalog or when the resources were originally published. We can generate a time series or animation of these additions, resulting in a timelapse view of the development of the catalog, both from the perspective of when the resources were created in the catalog and from when they were published.  

We can also use the web interface to step through year-by-year and see which areas of the catalog received the most contributions that year and which LCC classes were being added each year. Class P (Language and Literature) is the largest addition every year, but the time-lapse presentation emphasizes areas that received a high concentration of new records. The time lapse by date entered shows a sort of institutional view of what was processed each year, and perhaps certain priorities, while the time lapse by publication date is a more general indicator of the development and publication of materials throughout the history of the Library and its collections. 

LINK TO TIMELAPSE VISUALIZATIONS – Press the play button and use the buttons at the top right to switch between cataloged and publication date. 

Because we are mapping concepts and ideas into a two-dimensional space, we can also introduce other types of knowledge bases for comparison, not just catalog records. For example, we can show the overlap between a dataset like Wikipedia and the Library by vectorizing the 7 million English Wikipedia articles and overlaying how they fit with the Library’s vectors. This is possible if we use the same embedding model.  

Of course, it is not a one-to-one comparison. These are fundamentally different types of knowledge — a catalog record versus an encyclopedia article. But within the model’s latent space, you can get a sense of where there is overlap and where one system has more information than the other.  

In the visualization below, which projects the Library’s vectors together with English Wikipedia’s articles, we can see significant overlap and a lot of areas where one knowledge base has more information than the other. For example, Wikipedia has many individual articles about sports players, whereas the Library of Congress does not have thousands of individual player biographies.  

LINK TO WIKI X LC VISULAZATION 

We can also see certain artifacts of how a knowledge base is built, like the Wikipedia cluster of disambiguation pages that have no correlation to the Library’s knowledge base. This is just one example of how you can overlay disparate systems of knowledge and explore how they compare. The most obvious application, of course, would be to use this sort of analysis on other catalogs and see how they relate to each other. 

To conclude, this experiment suggests that the approach of vectorizing the descriptive metadata of catalog records can be used for various applications, and it works well due to the distinct nature of catalog records. Many other applications could be developed using the same methodology, which will likely improve as newer, more capable embedding models are released regularly by research groups and companies. 

Below you’ll find some technical information if you’re interested in the details of this experiment, as well as some download links if you’d like to try out the embeddings yourself. 

Technical Details 

• Corpus: Approximately 20.3 million bibliographic records from the Library of Congress catalog through the end of 2025. Only descriptive metadata was embedded (title, subject headings, classification, summary/table-of-contents notes, etc.).
• Embedding model: Amazon Titan Text Embeddings V2 (via AWS Bedrock), producing a 1,024-dimension vector per record.
• Vector DB: We used AWS S3 Vector bucket for storage and querying our 20M vectors. 
• Dimensionality reduction: PCA from 1,024 down to 50 dimensions, then Parametric UMAP down to 2D — a PyTorch multilayer perceptron (4 hidden layers × 512 units) trained for 20 epochs on Apple Silicon (MPS), with an added correlation-preservation loss term. Training used a stratified sample of approximately 5.4 million records (capped at 30,000 per LCC subclass so small classes get equal say in the layout). The trained network then transformed all 20.3 million records. 
• Rendering: Datashader aggregates all 20.3 million points per zoom level with logarithmic alpha shading, rendered into a standard XYZ tile pyramid (zoom 0–10, 256 px PNG tiles — about 620,000 tiles) via 2,048 px supertiles processed in parallel. 
• Viewer: Plain Leaflet with CRS.Simple. The entire visualization is comprised of static files (about 5.6 GB total). No server or database is required to host it. 
• Click-to-inspect: An adaptive quadtree shards the 20.3 million records into approximately 11,000 static JSON files; a click fetches only the leaf under the cursor to show class percentages, sample records, and links into the Library catalog. 
• Labels: Class labels are placed at density peaks of per-class 2D histograms (with connected-component centroids for sprawling regions) and revealed progressively by zoom. Language “watermark” labels are derived from the equivalent 008/041 fields. 
• Time lapse: Per-year frames rendered with Datashader and assembled with ffmpeg into MP4s. The web explorer version uses one WebP peak-frame per year with a single JSON stats file. 
• Stack: Python 3, NumPy / pandas / PyArrow, PyTorch, Datashader, Leaflet. The full pipeline runs on a single computer. The layout is trained in about 7.5 hours. 
• Embedding Files: https://id.loc.gov/download/#embeddings.

What is your job at The Library of Congress and what was your role in creating this collection? 

I am a Preservation Librarian in the Conservation Division. Part of my role is ensuring that conservators and preservation scientists have access to the resources they need for their work, including new scholarly publications. As more of those resources became available only online, I realized we needed to preserve access for the future, which led to creating the Preservation and Conservation Web Archive. I am the Collection Leader, and I’ve researched the field and worked with colleagues across the Conservation Division to identify resources to include in the collection. 

Briefly describe the collection and how you selected materials for the web archive. Did you have any specific goals or considerations in mind? 

The web archive was initially created to support the conservators and preservation scientists at the Library, and I consulted with them to identify websites and other online resources that were important to their work. In keeping with the Library’s role as a leader in conservation, the scope of the collection has since expanded beyond those immediate needs to include materials from around the world and about the conservation of all types of materials. I researched international preservation and conservation organizations to find sites to add to the collection and hope to continue expanding that aspect of the web archive in the future.  

We want to build a comprehensive collection of information related to preservation, ensuring that the vital research, methodologies, and other resources remain accessible despite the transient nature of the web. The archive includes preservation vendors, educational institutions, international nonprofits, and scientific resources. By consolidating these materials into an organized and easily accessible platform, the archive will support both current and future work across the field. 

What are some of the highlights? 

I’m happy that we have the Conservation and Art Materials Encyclopedia Online (CAMEO), which is maintained by the Museum of Fine Arts, Boston. It is a useful tool for understanding and identifying specific brand names or chemicals, for example. Having that archived means that conservators in fifty or a hundred years can identify exactly what a conservator today means when they describe a material or process in their treatment report.  Similarly, we’ve archived the website for the Image Permanence Institute, which conducts research on photographic preservation. The Library and conservators more broadly rely on its research to understand which materials are safe to use with photographs. 

We are also working to preserve the Iron Gall Ink Project, which documents an international, multi-institutional project to understand the science of and conservation needs of iron gall ink, the most common writing ink of the preindustrial era. The website has been online since 2005, which is unusually long for a resource of its kind, and will be added to the Library’s web archives in an upcoming release cycle. It is especially important to preserve websites like this because the people who maintain and advocate for them may retire or move to other institutions, and the resources they provide could disappear. 

What challenges have you faced? 

One of the biggest challenges has been identifying smaller, independent websites that may be valuable to the field but are harder to discover. As we broaden the collection beyond Anglophone book and paper conservation, it can be difficult to know which resources are important, or even what resources are out there. To help address that, I’ve been reaching out to the broader conservation community through professional conferences to learn what resources practitioners find most valuable. 

Why do you think web archiving is important for documenting this subject? How do you imagine researchers, now or in the future, might use this web archive? 

One specific example of why web archiving is so important is preserving the websites of library and conservation vendors. The Library of Congress has a rich collection of vendor catalogs dating back to the late nineteenth century, but as print catalogs disappeared in the early 2000s, we weren’t able to add to it. The web archive preserves the online presence of vendors and extends that historical record. Researchers interested in how vendor products, services, and practices have changed over time can now trace those changes across both the print and digital eras. 

In 1851, more than a decade before the Emancipation Proclamation was issued, the School for Colored Girls opened in Washington, D.C. with the mission of training Black women educators. The institution was the first of its kind in the District and only the fourth in the nation. Today, the Library of Congress and the By the People volunteer transcription program invite you to celebrate the life and legacy of its courageous and barrier-breaking founder, Myrtilla Miner (1815-1864).

Miner spent her young adulthood teaching in New York, Rhode Island, and Mississippi. In the latter she was denied permission to conduct classes for enslaved children, after which she resolved to open her own school for “colored” women teachers. 

Frederick Douglass recalled that when she came to him asking for support for her endeavor, he “thought the proposition was reckless, almost to the point of madness” but “her resolution was taken, and was not to be shaken or changed.”  

Indeed, Miner persisted despite the danger. In May 1860, she wrote of waking to find her house had been set on fire by “an incendiary.” She wrote, “had I slept five minutes more it would have been too late – for in that time it was in my chamber, within two yards of my bed.” In the same letter, Miner alludes to previous harassment: “For years I had expected this, so it did not take me unawares, but for the last year the ruffians had been so quiet, I tho’t they had given me up.” 

Douglass later wrote of Miner, “when I remember how low the estimation in which colored people were then held, how little sympathy there was with any effort to dispel their ignorance, diminish their hardships, alleviate their suffering, or soften their misfortunes,  I marvel all the more at the thought, the zeal, the faith, and the courage of Myrtilla Miner in daring to be the pioneer of such a movement for education here, in the District of Columbia, the very citadel of slavery, the place most zealously watched and guarded by the slave power. . .” 

Miner’s success was even more wonderous given she suffered from periods of illness throughout her life. In 1857, she left teaching and the management of the Normal School due to her poor health. She moved to California in 1861 to try to improve and, there, was tragically injured in a carriage accident, eventually leading to her death. She died in Washington, D.C. on Dec 17, 1864. 

The Myrtilla Miner Papers, held by the Manuscript Division of the Library of Congress, include her own correspondence and notes, subject files on the Normal School, including its founding, operations, and students, and an unpublished biography of her life. The collection documents her work to create access to education for and by African Americans and highlights the role of women in that effort, as well as the larger abolition movement in the 19th century. 

You can become part of the legacy of abolitionist and educator, Myrtilla Miner, by transcribing her papers with By the People. The Library of Congress transcription program invites you to become a virtual volunteer by deciphering the text of historic documents on the By the People website. Your work will improve the visibility of Miner’s story along with the women she educated. Completed volunteer transcriptions go back into digital collections on loc.gov to enable page-level, keyword search of digital collections and improve the accessibility of these texts. We would love to hear what you find in her papers! 

At the beginning of July, I got to travel to Wyoming to do some in-person programming with our library collaborators in Cheyenne and Laramie!

Cheyenne

In Cheyenne, because so many locations are within walking distance from the library, we did a walking tour of three Anywhere Adventures stops.

The first stop was at the Wyoming State Capitol building to see the statue of Esther Hobart Morris. Morris is known for being the first woman to hold public office as justice of the peace. We chose this first stop to ask some ice-breaker questions to get to know one another.  Participants had to choose one of the following difficult questions, and defend their answers:

• If you could only eat one food for the rest of your life, what would it be?
• What’s the ideal breakfast?
• Are there more doors or wheels in the entire world?
• If everybody had to read one book, what would you assign?

Our second stop was at the Historic Atlas Theatre, where theater staff and actors were busy getting ready for their opening night production of The Melodrama!, a comedic play fashioned after melodramas of the Old West. The play encourages audience participation and is part of a summertime tradition in Cheyenne going back decades. We got to visit the inside of the theatre (and even climbed backstage) to see its historic hand-painted curtain.

Our final stop was at the Wyoming State Museum, where we rang a small bell for good fortune and took a texture impression of the Liberty Bell replica from the 1950 Independence Bond Drive.

Laramie

In Laramie, the Albany County Public Library (ACPL) set up a table outside of Bond’s Brewing during the city’s annual Jubilee Days, a weeklong celebration marking Wyoming’s statehood. Folks passing by could make a button of Anywhere Adventures themes, pick up an Adventure Log, and grab some materials about doing research locally or online with the Library of Congress.

We also had a printed copy of all the Laramie Sanborn maps from 1923. The fire insurance maps were created by the Sanborn Map Company. They give readers an idea of how their neighborhoods and cities developed over time. Participants could flip through the maps to find their homes.

Southeast Wyoming

I also got to visit some of the Anywhere Adventures Southeast Wyoming locations. It was cool to see these places in person and to stumble upon things that I didn’t even know existed, like the irrigation channels created by the Civilian Conservation Corps at the High Plains Arboretum.

Collages of different locations in SE Wyoming created by Vivian Li on the Anywhere Adventures website.

These collages include photos I took at each of the Anywhere Adventure locations, including the outside of the Atlas theater, the cottonwood trees lining the Tree Walk at Lion’s Park, the apple orchards at the High Plains Arboretum, close ups of the Esther Morris statue, and the Liberty Bell replica outside the Wyoming State Museum.

Southeastern Wyoming Travel Log: 

Download your own travel log for Southeastern Wyoming. Learn more about the Anywhere Adventures locations and record your thoughts and ideas! Follow Vivian’s instructions to assemble your log after printing.

Coloring Pages: 

Get creative! Download and print out coloring pages designed by Vivian and inspired by Anywhere Adventures locations in Southeast Wyoming.

This is the last blog post in the Anywhere Adventures blog series. Head over to the Anywhere Adventure’s experiment page to see all of Vivian’s past work from her residency, read other blog posts in the series, and explore her mobile-first website, Anywhere Adventures. 

The following is a guest post by the 2025-2026 Innovator in Residence, Vivian Li. As part of her work with the Library, Vivian hosted a series of community outreach events, including zine-making workshops in partnership with the Chicago Public Library. Read on to learn how you can also make a zine and check out the end of the post to print your own copy of research zines created by Vivan about various LOC collections.

In Chicago, the Anywhere Adventures team ran a series of workshops inviting people to make zines about their neighborhood. These zines could describe a piece of history that happened on your block, a memory you have, some kind of guide, or anything you want to share and preserve about your piece of the world.

I’ve been copying and gathering these zines into a community artbook called the Zine Atlas. If you would like to make a zine and contribute it to the Atlas, or if you just want to participate in the writing prompts we used in Chicago, you can do it yourself at home!

Materials needed:
– Paper (2 sheets)
– Scissors
– Drawing supplies (just a pencil/pen is fine, but feel free to use anything!)

How to fold a magic zine

Magic zines are magic because they turn a single piece of paper into an 8-page book without any staples, glue, or anything to bind them. You just need some clever folding + cutting.

Instructions

1. Fold a sheet of paper in half, lengthwise.
2. Open the sheet of paper.
3. Fold the same sheet of paper in half, widthwise.
4. Open the sheet of paper. Your sheet should have a vertical and horizontal crease at the center.
5. Fold the right side of the sheet and align the right edge with the vertical crease in the middle of the sheet of paper.
6. Open the sheet of paper.
7. Fold the left side of the sheet of paper and align the left edge with the vertical crease in the middle of the sheet of paper.
8. Open the sheet of paper. You should now have one horizontal crease, and three vertical creases on the sheet of paper.
9. Fold the sheet of paper in half, widthwise.
10. Take your scissors and cut from the folded edge along the horizontal crease towards the center. Make sure you DO NOT cut past the vertical crease in the center.
11. Open the sheet of paper. There should be an opening in the center of the sheet from where you made your cuts and two slits at the left and right of the opening.
12. Fold your sheet in half, lengthwise. The opening in the middle of the sheet will flare. Push both sides of the paper together so they meet in the middle. The sheet will now look like a plus sign.
13. Fold the sheets onto each other by bending at the creases.
14. Set this aside for now as you brainstorm below!

Brainstorm (optional)

To help you start brainstorming, think about these main questions: What neighborhood are you writing about? Is it the one you currently live in, or one that you lived in before?

Take your second sheet of paper and cut it into four quarters. You’ll use these quarters to write down your responses to each of the prompts below. You’ll only need to use three of the quarters, so feel free to use the last one for another project.

Prompt #1: Make a List

On the 1^st  quarter, start a list of all the words you associate with that neighborhood. They could be factual words, like the streets that cross through it, or the landmarks in the region. They could also be descriptive words, including the smells you might experience there, or the feelings you have when you arrive.

Set a four-minute timer and write down all the words you can think of!

Prompt #2: Draw a Map

On the 2^nd  quarter, create a map of your neighborhood. It can be a map at any scale. For example, at the scale of the world, where would your neighborhood be? Or it could be really zoomed in — just one corner of your backyard.

Some things you could add to your map:

• Points of interest – If someone found themselves in this area, what is something they should definitely stop and see?
• Orienting landmarks – How would someone establish where they are? What items in this area don’t change?

Prompt #3: Draw an Item

On the 3^rd quarter, draw an item or object that you associate with your neighborhood. It’s okay if your drawing is bad! See what you can make, and as you’re drawing, you can think to yourself:

• Is the item or object large? Small? Heavy? Light-weight?
• Who interacts with this item or object?

Make your zine

Make your zine about your neighborhood! From your brainstorming, you could decide to focus on one specific portion – or maybe it can be a general overview. There are no right or wrong approaches.

Some things you could include in your zine:

• Text or other words
• Photos
• Lists
• Maps
• Drawings

You could also paste in parts of your brainstorming sheet, if you’d like!

Find your historic map

The Library of Congress has so many digitized maps available online for free at loc.gov. Search your town or neighborhood to see what you can find.

You can print out your map and annotate it. Ask: How have things changed? What has stayed the same? You could also add any drawings or writings related to your zine on top of the printed map.

Share

A big part of zine-making is sharing your zine! Make copies of your zine and give them out to people in your life. If you’d like to contribute your neighborhood zine to the Zine Atlas, you can scan and send it to: lc-labs@loc.gov.

In your email, please include:

• The location your zine is about, at any scale; it could be an intersection, town, state, etc.
• Attribution for the zine – how you’d like to be credited for your work.

And if you don’t know where to find a copier or scanner, you could try visiting your local library!

As an end to her residency, Vivian created a zine series about how to research 7 different Library of Congress collections: Sanborn Maps, Historic American Buildings Survey, the Historic American Engineering Record, the Historic American Landscapes Survey, the Carol Highsmith Archive, Chronicling America, and the Occupational Folks Life Project. They’re all available for you to read, and print on the Anywhere Adventures Experiment page.

Leah: Hi Sylvia! It’s great to have you on our team this summer. Could you tell us about your role at the Library, and what you’ve worked on?

Sylvia: My role at the Library this summer was Artificial Intelligence Community Specialist Intern, which involved working with the Digital Strategy Directorate on various AI literacy, strategy, and community engagement initiatives.

One of the main AI literacy projects I completed was taking the Library’s Principles for Responsible AI Use and mapping them to actionable AI-related competencies. I shared this in a chart format that was easily readable and accessible for Library staff members regardless of their familiarity with AI, hopefully to help empower them to make informed judgements about the technology. I also conducted a survey with my fellow interns to gauge their feelings about AI to help inform the direction of AI literacy at the Library, and was involved with AI strategy development.

Leah: Tell us a bit about your background and journey. What professional or educational experiences led you to apply for this internship?

Sylvia: My main career interest is in contributing to the responsible and ethical development and implementation of artificial intelligence, so this internship was perfect for me! I have a BA in Computer Science and am currently finishing my MA in AI, Ethics and Society, which I was encouraged to do after my experience working with ethical AI in education. I was looking for hands-on experience in AI governance, ideally in a public service-oriented institution, so I was excited to see this internship opportunity open at the Library.

Leah: What in your background, coursework, or job experience has been useful for your work on this project?

Sylvia: The work I’ve done in my master’s program has definitely informed my work here at the Library. During my program, I learned how to evaluate past and emerging AI systems with a critical lens regarding not only their efficacy but also their ethics. Having a strong sense of the factors involved with ethical technology is important when we are thinking about AI at the Library, because all technology implementation here needs to be equitable, sensitive to privacy concerns, and human rights-oriented.

My master’s program also helped me have a more nuanced understanding of the current AI landscape, including how other organizations are approaching AI adoption and resistance, which was also invaluable context. Otherwise, I also found my past experience working in a start-up environment to be helpful for my internship at the Library. That fast-paced work environment taught me how to work independently, take initiative, and communicate clearly, which are great skills for any work environment.

Leah: Is there anything about working on this project that you found surprising or unexpected?

Sylvia: This was my first time working for a government agency, so I felt like I was consistently learning new ways of approaching AI issues, including the specific considerations that need to be made for government work. Especially since AI is developing so rapidly, the considerations that need to be balanced, especially for a cultural heritage institution, are unique. It helped me take on a new point of view that helped me develop into a more well-rounded AI governance professional.

Leah: Tell us a bit about how this work connects with your career goals. How has this experience helped you to further develop or refine your goals and interests?

Sylvia: This internship experience absolutely reaffirmed my career goal to work in responsible AI, and I hope to use this experience to get into a similar role. It also piqued my interest in working with cultural heritage institutions, both because of how wonderful the people and initiatives are and due to the unique challenges these institutions face in the age of AI. This internship was perfectly aligned with my career goals, and I am so lucky to have gotten to experience it! I also admire the Library for hosting an internship like this, which will undoubtedly help positively shape the next generation of young professionals in AI governance.

Leah: Based on your experience in this internship, are there any things you are going to focus on getting out of your next educational experience? Has this experience shaped your ideas about your future work and educational program?

Sylvia: As I am almost finished with my master’s program, my next step is looking for a full-time role. This internship has helped significantly with the direction of my job search through giving me a clearer sense of the actual work I would like to be doing on a day-to-day basis. The experiences I had and skills I developed at the Library make me feel confident to pursue roles in the government and cultural heritage institutions, but I also feel like they are very transferrable to the private sector as well. Overall, I feel like this experience has prepared me well for my next role.

Reference staff across the Library include web archive collections in their Online Research Guides right alongside digitized historical documents, e-books, databases, and more. Which guides created by the Library’s subject experts feature web archive collections? If you’re an educator, get inspired by these examples. And if you’re a researcher at any stage, start clicking. The range of Research Guide subjects is as deep as the web itself.

African, Jewish, and Middle Eastern Studies: Digital Resources at the Library of Congress

This research guide provides a compilation of primary and secondary source materials, reference works, free and subscription databases, recorded programs and more, selected by curators in the Library’s African and Middle Eastern Division. Web archive collections are listed under Digital Collections & Exhibitions. The section includes the African Government Web Archive and the Crisis in Darfur 2006 Web Archive.

Bosnia & Herzegovina and the Bosnian Collections in the Library of Congress

Bosnian materials are located throughout the Library of Congress with the European Reading Room as the main point of contact. This overview of the collections describes books, electronic resources, journals, visual materials, and other special formats. The geographically-focused guide from the European Reading Room includes a Web Archiving subsection under Digital Resources with the East European Ministries Web Archive and the Bosnian Political and Social Issues Web Archive. The East European Ministries Web Archive is also featured right on the guide’s homepage!

Business and Labor History: Primary Sources at the Library of Congress

This guide highlights those special collections at the Library that would be of interest to researchers studying business and labor history with an emphasis on personal and organizations papers and related resources. Reference librarians from the Science and Business Reading Room and the Manuscript Reading Room collaborated on this guide which includes the Economics Blogs Web Archive and Business in America Web Archive in the A-Z List of Primary Sources section.

Comics and Comic Books: A Research Guide

This is a guide to using the Library of Congress’ Comic Book Collection, as well as general information on comics research. Staff from the Newspaper and Current Periodical Reading Room nest Web Archives under Digital Collections. Three featured web archive collections are Small Press Expo Comic and Comic Art Web Archive, Comics Literature and Criticism Web Archive, and Webcomics Web Archive.

Guide to Law Online: U.S. Federal

This guide, prepared by the Law Library of Congress, includes links to free online legal resources regarding the United States of America, focusing on its constitution; executive, legislative, and judicial branches; legal guides; and general sources. The Law Library of Congress includes two of their web archive collections in this research guide: Executive Branch Federal Government Web Archive and Legal Blawgs Web Archive.

South Asian Collection: Asian Collections at the Library of Congress

The Library’s South Asian collection comprises thousands of titles in over 100 languages. This guide provides an overview of these materials, search strategies, and how to access digital collections, e-resources, and web archives. Check out the Web Archives section of this guide created by reference staff in the Asian Reading Room. It features twelve (!) web archive collections from the Library, as well as related web archive collections from other institutions.

Performing Arts Web Archives at the Library of Congress

This was the first Online Research Guide at the Library to solely feature web archives. Explore web archive collections maintained by the Music Division and collections with related content from other areas. This guide describes digital collections of web archives related to the performing arts, how to navigate the collections, potential research questions, and how web archives relate to Music Division holdings.

A postscript to this rich sample of LC Research Guides is that our web archive collections have been featured in ones created by other institutions. The September 11, 2001 Web Archive has been particularly referenced as a source for contemporary history. For example, this collection has been listed in the 9/11 Resource Guide from Illinois Wesleyan University Lewis A. Jackson Library, in Historical Primary Sources: United States: Primary Sources: 21st c. by the University of California at Berkeley Library, and in History: American History by Peru State College Library. The web archive’s inclusion in these reference resources gives even more context and importance for classroom instruction. For example, the September 11, 2001 Web Archive is part of undergraduate history course syllabi at the University of Dubuque (HIST122: World Civilization II: Digital Archives) and Sam Houston State University College of Humanities and Social Sciences (HIST 3399: Remembering September 11, 2001). Even the Connecticut State Library lists this collection among its History Day resources (History Day – Revolution, Reaction, Reform: September 11, 2001).

The Web Archiving Program is grateful to our colleagues from across the Library and other institutions for their collection advocacy work to include the web archives in online reference materials. While this blog post is by no means exhaustive – the web and the Library’s web resources are constantly updated – I hope this gives all curious minds places to start exploring websites as primary sources.

As opportunities to discover uncharted patterns in our nation’s history expand with artificial intelligence-powered research tools, technical approaches that promote authenticity and context are needed more than ever.

Recently, the Library’s Digital Strategy Directorate teamed up with Arizona State University Cloud Innovation Center students to investigate how technical innovation can be grounded in truth using founding era documents.

The Arizona State University Artificial Intelligence Cloud Innovation Center (ASU AI CIC) brought their technical expertise, design thinking, and rapid prototyping to the experiment, one of several activities the Library has undertaken to celebrate America 250. The challenge, developed by Library staff members and ASU AI CIC student workers skilled in technical architecture and cloud services, was to develop an approach and prototype a tool to would enable contextual research in a broad collection of documents from America’s founding.

The History Reimagined: Library of Congress challenge was, as the project team described it, an “experiment for an innovative AI-driven solution that transforms public access to foundational U.S. historical documents.”

They added: “This initiative aims to expand access with new approaches for users to search, interpret, and engage with complex materials like the U.S. Constitution by enabling natural language interaction and providing deep, contextual understanding grounded in the Library’s vast collections.”

The goal was to build a tool using AI to help discover historical documents from 1760 to 1810. These documents included materials about the Constitution’s history, congressional legislation, historic newspapers, and related legal materials.

The team of ASU student workers included Ridham Sonani and Hemanth Koneti, backend developers and cloud engineers, and user interface/user experience designer Ashik Tharakan. As Sonani described the collaboration’s goal, “We aimed to help researchers, students, and policy analysts explore early American history by building a[n] AI tool that can answer questions about congressional bills from 1789-1875 and historical newspapers from 1770-1810.”

The team used the Library’s application programming interfaces (APIs) as well as AI-ready data processed for another America 250 project, Revolution Crossroads (a collaboration between the Library and the Smithsonian Institution), to procure the data necessary to understand entities, map relationships, and power-queries. The Congress.gov API surfaced bills from the first through 16th Congresses (1789-1875), while the Revolution Crossroads project data provided historic newspapers from the Chronicling America Historic American Newspapers collection (1770-1810).

Determining which data would provide a source of truth for the prototype was only the first challenge. The team also explored how to build an AI-powered tool to use the available materials. Approaches including GraphRAG processing, a type of Retrieval Augmented Generation (RAG), enabled the extraction of entities (like people, places, and concepts) through text search and the tracking of relationships using knowledge graphs. This approach enhanced the search capabilities of the tool while maintaining its foundation in authentic documents and data.

“In particular,” said Sonani, “using Neptune Analytics to create a knowledge graph instead of just vector search was a turning point; it let the chatbot connect related documents across time periods.”

Connecting to a knowledge base in this way could allow anyone with an interest in history to make connections across huge swaths of documentary evidence. According to Koneti, the team’s decisions about the tool’s architecture allowed use of the documents to “build knowledge graphs that connected historical events, legislation, and people beyond traditional vector search.”

Although the tool was produced as a prototype, this approach provides the Library with insights into how AI could help future historians understand those connections.

Because the Library serves Congress and the American people, the ASU and Digital Strategy Directorate team also considered how user needs might differ depending on role and experience. Agreeing that providing different paths would refine the search experience, the team created user personas to customize response styles.

“I particularly enjoyed designing the persona selection feature, which allows users to receive information in a style that matches their needs,” said Tharakan. The work, he went on, allowed the team to “think through the entire user journey, from selecting a persona to asking questions and reviewing cited sources, while ensuring the experience feels smooth, trustworthy, responsive, and easy to use.”

In addition to the knowledge base and historical coverage, relationship discovery, and multiple personas for different user expertise levels, the team took a comprehensive look at both sources and user experience. This allowed them to incorporate source citations with links to original documents and session memory for contextual conversation continuity.

Complete information about the Cultural Heritage Chatbot – Library of Congress can be found in the LOC Project GitHub repository. It includes an overview of the project’s aims and technical approach; an explanation of the personas; architectural diagrams; the Amazon Web Services (AWS) architecture for the frontend, backend, AI layer, data storage, memory, and data collection; diagrams of data flow; example queries; and credits. The Architecture Deep Dive and Deployment Guide allow anyone to experiment with these technologies and approaches.

Perhaps most importantly, the team managed both the data and cutting-edge AI technologies to ensure this project maintained its grounding in authentic documentary evidence, avoided hallucinations, and provided truthful information and analysis.

In an era of increasing data proliferation data and evolving AI-powered tools, this commitment to truth and authenticity is a laudable aim.

A Thank You From Us 

Thanks to everyone who came out to The Library of Congress National Book Festival in August! “New Online at the Library of Congress” newsletter contributors engaged with hundreds of attendees about virtual volunteering, web archiving, and Library of Congress digital collections. We hope to see you again next year! In the meantime, we hope you’ll enjoy learning about the latest additions to the Library’s digital collections: 

United States Supreme Court Records and Briefs 

The Law Library of Congress Records and Briefs collection consists of the briefs and transcripts of records submitted to the Supreme Court of the United States on appeal from federal courts and state appellate courts. This initial release includes nearly 10,000 documents from 3,000 cases spanning the years 1999-2013. The collection offers full-text searching as well as metadata including case name, U.S. Reports citation, and file date. 

National Press Club Speeches

Since 1932, the National Press Club has hosted luncheon gatherings allowing presidents, visiting world leaders, and others to address the press and answer questions about current affairs. In 1969, the Press Club donated audiotapes of nearly 2,000 talks recorded since 1952 to the Library of Congress. 62 of these talks are now available online, including one featuring the Library’s own James Billington in 1989 announcing the American Memory project, which was created to make digitized versions of selected materials more accessible nationwide. 

Ethiopic Manuscript Collection

This Ethiopic manuscript collection features rare materials at the Library of Congress related to the Horn of Africa – home to one of the oldest African manuscript traditions – and is made available online through the generous support of the Lilly Endowment, Inc. The Library’s collection contains over 250 materials dating from the early modern period to the present.  

The vast majority are written in Ethiopic languages, primarily Ge’ez as well as some in Amharic and Tigrinya, with some representation of Arabic. Many of these centuries-old codices are brilliantly illuminated, and include Gospels, Psalters, hagiographies, mystical treatises, amulet scrolls, hymnals, prayer books, collections of sacred poetry and handbooks on the practice of folk magic, as well as rare medicinal books and manuscripts on the history of the Horn of Africa. These volumes illuminate the numerous religious cultures of the region, its literature, sacred art, and scribal traditions. Digitization of the collection continues and additional materials will be added to this presentation as they become available. 

Now online: two new Community Collections Grant recipients 

Community Collections Grants from the American Folklife Center supported contemporary cultural field research from 2022-2024. Read more about the program on loc.gov. 

California’s Indigenous Maritime Traditions

This collection is based on the 2023 Community Collections Grant project, “America’s First Boats and their Makers: Securing knowledge of the Kelp Highway and California’s enduring, sustainable Indigenous maritime traditions.” In collaboration with the University of Southern California, the project was led by Los Angeles Indigenous community members whose maritime coastal lifeways depended on human powered vessels, ti’ats (tee-ahtz), which people paddled while settling and living along the Pacific coastline and its adjacent islands. The collection comprises fourteen interviews with Ti’at Society members, as well as photographs of events Ti’at Society organized or participated in. 

La Musique Nous Réunit: Documenting Louisiana French Music

Supported by a 2024 Community Collections Grant (CCG) from the American Folklife Center, the New Orleans Foundation for Francophone Cultures (Nous) documented oral histories and performances by six prominent Louisiana musicians and musical ensembles. The Nous team, led by Scott Tilton and Rudy Bazenet with artistic directors Katya Vaz and Autumn Palen and film director Cory St. Ewart, conducted and filmed a series of semi-structured interviews with French-speaking, Louisiana-based musicians and bands spanning different genres such as Cajun music, zydeco, and swamp pop. 

New foreign legal gazettes, oral histories, and more

We are happy to share a few other digital collections updates. 1,400 issues of Boletín oficial del estado from Spain have been added to the Foreign Legal Gazettes in the Law Library of Congress online collection. And the Benjamin Luft Collection of 9-11 Oral Histories collection has been updated to incorporate individual interviews and related collection materials. This collection features 484 interviews of responders to the September 11, 2001 terrorist attack on the World Trade Center, and some of their family members, as well as photographs and manuscript materials documenting their experiences at the site. 

World War II-era government contracts have also been added to the digital collections on loc.gov. Initially highlighted in a 2013 Inside Adams blog post and in the Federal Government Contracting: A Resource Guide from the Science & Business Reading Room, most of these sets were published by the United States government. Each contains a list of public contracts information by company and largely date to World War II.  

New volunteer transcriptions and datasets

Additional volunteer transcriptions of Theodore Roosevelt’s Papers are now integrated back with their original digital collections on loc.gov, improving search and discovery for all. Completed by participants in the By the People crowdsourced transcription program, these transcriptions of Roosevelt’s papers date to 1914-1915, right at the outbreak of World War I in Europe.  

The Louis Malesherbes Goldsborough Papers transcription campaign dataset is also now available in the Selected Datasets Collection on loc.gov. 

And have you ever wanted to search across all 30+ By the People crowdsourced transcription datasets? Now’s your chance! This month, we published the first-ever iteration of the “By the People comprehensive transcription data package.” The dataset combines full text transcriptions from contributors, tags from contributors, selected metadata fields from Library of Congress staff, and URL pointers to download images in different formats. A guide to this data can be found on the Library’s GitHub. 

Learn more about By the People and how to get involved in our video below or on YouTube. 

New onsite-only digital content

New items are added every week into stacks.loc.gov  – the Library’s primary onsite platform for accessing restricted digital content. To learn more about Stacks, check out this video from our team:  Access the Digital Stacks On-Site at the Library of Congress!

Recent highlights from Stacks include The Lost Lodge gold, Justin Jefferson vs. Jerry Rice: football legends face off, Heart rate variability: science and strategies for peak performance, 10-minute true haunted house stories, and Dark faces in cyber places: shattering stereotypes and leading with legacy in cybersecurity. 

And some seasonal additions to Stacks include: NFL football: a history of America’s new national pastime, Ultimate college football road trip, Twelve days of autumn, Summer green to autumn gold: uncovering leaves’ hidden colors, Betty the Yeti and the first day of school. 

Please reach out to a librarian at ask.loc.gov with questions about accessing these materials using Stacks. 

Library and Information Science web archive now available

The Library and Information Science Web Archive is now available on loc.gov. The collection includes sites focusing on library and information science, library organization sites, and library-related blogs representing a snapshot on the state of library and information science in the United States. Sites include association blogs, library science education sites, archives and special collection sites, and more. 

Learn more about the Library’s web archives in this Signal blog post from 2026 Junior Fellow, Nicky Martin: What Can a Computer See?: Testing the Limits of Automated QA for Web Archives. 

This online presentation will share about the life and legacy of the 20th president, explain the history of the Garfield Papers at the Library of Congress, and show how you can get involved with volunteer transcription for his official papers. 

Since 2021, By the People volunteers have completed transcriptions for over 50,000 pages of Garfield’s papers. Transcriptions created by virtual volunteers are published back on loc.gov alongside their original digital collections, and greatly improve readability, discoverability, and accessibility for all. Many of these volunteer transcriptions of Garfield’s papers are available now – you can explore them here. 

We’re also excited to announce the opening of the next batch of Garfield’s presidential papers for volunteer transcription! This new transcription project, General Correspondence, Vols. 95-108, October-November 1880, spans the month of his election as President in 1880. His election “opened a floodgate of correspondence from friends and associates, well-wishers from the general public, and those who sought to curry favor with the president-elect.” You can dive into the 10,000 new pages available here.

We look forward to seeing you at the webinar so that you can join in on transcribing this most recent release of Garfield presidential papers! 

CLICK HERE to register for the Sept. 24 event on Zoom