We began a project in 2025 called "DigitAI" (named for our "Digit" program at Penn State Behrend) with intentions to make a "black box" technology explainable to ourselves and to find out if we can design a minimal and efficient small-scale language model system that can run on a laptop computer. At first, DigitAI was an experiment to supply a Small Language Model (SLM) with Retrieval Augmented Generation (RAG)–giving it query access to a structured “ground truth” knowledge base derived from the XML structure of the TEI P5 Guidelines.

We launched our project with these primarily educational goals in view:

1. To learn how to design a small-scale “explainable AI” system, that is, a system designed to give students and scholars in arts and humanities fields their own agency and purpose in building, training, augmenting, and applying an AI with local installations.
2. To create a local coding assistant to serve as a guide for decision-making in humanities text encoding with TEI XML, or on a particular XML project with a defined encoding schema, responding to experiments by Khemakhem (2024) Scholger et al. (2024), and Pollin et. al. (2025).

Preparing our RAG involved parsing the TEI Guidelines (available as a complete XML document in P5.xml from the TEI Guidelines repository) in a context-aware manner, stored and accessed in information "chunks" based on the text content of XML element nodes.

At first, we tried to model the entire TEI Guidelines (available from the TEI GitHub repository in a single document as P5.xml) in a context-aware “chunking” strategy, mapping XML node data into a graph structure. Preparing the RAG resource (shared at the 2025 TEI Conference) involved a complex XSLT transformation of the TEI P5.xml document into neo4j's graph structure–attempting to convert every possible relationship of the paragraphs of the Guidelines chapters and their associated schema definitions. Apparently we were creating a bloated” and suboptimal version of the TEI, and worse, it appeared that our model was relying on word embeddings to translate the information it received from the incomplete graph RAG. We called a halt to the neo4J RAG database effort as impractical as soon as we learned about Model Context Protocol.

Model Context Protocol (MCP), introduced by Anthropic and now open source, helped us to concentrate on supplying XML technologies to the language model (e.g. Skywork AI 2025). We have now designed a new SLM with agency to apply MCP as a tool to address a TEI XML tree with XPath, XQuery, and XSLT. This time our DigitAI can more reliably address XML project files to explore a TEI project codebase and aid in precisely identifying and continuing partially completed encoding.

We hope to share a look "under the hood" at our SLM + MCP system at the 2026 TEI conference. As this is an area of expertise in our team and program, we are confident that if we can teach humans to write XPath, we can script what is needed for an agentic AI system to access an XPath MCP service.