Hunt Institute for Botanical Documentation
A Research Division of Carnegie Mellon University

Hunt Institute Archives Text Discovery Platform

Search a large and growing portion of our online collections, including handwritten documents.
PROTOTYPE

This prototype uses state-of-the-art artificial intelligence, including a vision-language model (VLM) capable of reading handwritten documents as well as typed and printed text, to create searchable transcriptions of digitized materials in the Hunt Institute Archives. This makes it possible to search the textual contents of individual pages, including material that may not be described in the archival catalog.

Use Keyword search for specific words, names, dates, scientific names, or phrases. Try Semantic search (experimental) to describe a topic, question, or kind of material when you do not know the exact wording used in the documents.

About the AI-generated transcriptions

The transcriptions are generated automatically from page images and may contain errors, especially with difficult handwriting, unusual names, multiple languages, image-quality problems, or complex layouts. They are intended primarily as a discovery aid rather than authoritative transcriptions.

Each result provides the generated transcription and links to the original digitized material and associated archival description so that readings can be checked against the source. The transcription workflow uses AI models run locally by the Hunt Institute.

About Keyword and Semantic search

Keyword search is the default and matches the wording in the transcriptions. Results contain all your terms. Use quotes for an exact phrase. Substring matching is supported, so aceae can find plant-family names ending in -aceae.

Semantic search (experimental) is useful when you know what kind of material you are looking for but do not know the words used in the documents. It ranks transcribed passages by similarity of meaning, so relevant results may not contain the exact words in your query.

Semantic queries can be broad research topics, descriptions of activities or relationships, or natural-language questions. For example:

Semantic search is not a chatbot: a question is used as a search query, and the system returns archival passages that appear conceptually related to it rather than generating an answer or summary. Short descriptions and ordinary research questions generally work better than lists of disconnected keywords. Quotation marks have no special meaning in Semantic mode. Cross-language matching may work in some cases, but it should not be treated as translation.

Keyword and Semantic search are complementary. Keyword search lets you require particular wording; Semantic search can surface differently worded passages about the same subject. Depending on the research question, trying both can reveal different useful material.

Open a result: use the prominent page-and-transcription link to see the metadata, PDF, and full transcription. Keyword-search terms are highlighted in the transcription.

Archives Collections Database (ArchivesSpace): the Collection, Item/Folder, and Digital Object links open the corresponding archival records. Collection-level dates describe the collection as a whole, not necessarily the specific item or page.

If a PDF does not load: on the detail page, use the Digital Object link, click “Go to file” in ArchivesSpace, and navigate to the page number shown here.

Current limitations
  • Automated transcriptions can contain missing or incorrect text or unintended repetition. Difficult handwriting, image quality, unusual layouts, and multiple languages can reduce accuracy. Always consult the original page image when an exact reading matters.
  • Semantic search remains experimental. Its rankings are an additional discovery aid, not a complete or definitive set of relevant results, and highly ranked pages can sometimes be only broadly related.
  • This is an active prototype. Search coverage, transcriptions, functionality, and the interface may continue to change as additional archival material is processed and the system is improved.

← Back to results

Page 10 · DO #4670 · 231_Rogers_Bx2FF31_r

Collection
David James Rogers (1918–2007) papers
Item/Folder
Student Taximetric Papers, Spring 1967
Digital Object
DO #4670, page 10
Collection-level dates
1948–1977
Open PDF at page 10 ↗

Page transcription

Page 6

APPROACH
The first step in an approach to any problem should be to define the
problem, determine what is of particular interest about the problem, and to
check ones premises in regard to the discipline or disciplines involved.
The problem herein referred to was a group of plants collected by my professor in Taximetrics, purported to be representatives of the Family leguminosae,
Genus Astragalus. So it seemed they were. My objective was to be a classification of a number of these plants by the taximetical method I was to
learn in class.
Basically, the plants interested me because the genus contains species that are poisonous range forage plants quite unhealthy for cattle, and are
widely distributed problem for ranchers in Colorado. This is primarily because the genus contains over 300 species with much intergradation, which
makes it very difficult to tell the poisonous species from the not so, or non-
poisonous ones. My first steps were to consult respected authorities on the
genus such as "The Atlas of American Astragalus" by Barneby and "The Manual of Colorado Plants" by Harrington (containing a key and descriptions by Porter).
These works gave me some insight into what characters are important (i.e., help to classify) for the genus. There were of course many such characters, but I
selected only 20 (those I felt would be easiest to collect) to keep the study within the scope of the class. Also from my readings in Barneby I discovered
that historically the genus had been classified in two ways: one based primarily on gross morphology and a second based primarily on floral morphology.
It struck me that it might be interesting to contrast these two ways of classifying the genus to determine if there would be a difference in the results. I
hypothesized that given the assumptions of "biological order," a common genetic make-up, and reasonable care in selection and measurement of characters, the