Social Work Meta-Data

Empirical demonstrations · run locally, published openly

Demonstrations

The claims this project makes — that small, free, local AI models can query and search these databases — are tested, not asserted. Each report below is a complete, reproducible study run on a single MacBook Pro (Apple M5), with every result traceable to the code and raw outputs in the repository.

REPORT 01 · TEXT-TO-SQL

Can small local models query the databases?

Open-weight models — given nothing but the project's published skill files — answer 48 research questions across six task categories, from simple counts to multi-step analytical queries. Their SQL runs verbatim against the live databases; every answer is scored against pre-registered references. Includes the full question-by-question record with every model's SQL, and the skill-hardening feedback loop the first round produced.

Read the report
REPORT 02 · WORD EMBEDDINGS

Which embedding model best searches social work literature?

The benchmark behind this project's semantic search: embedding models — open-weight and commercial — compared head-to-head on retrieval over tens of thousands of social work abstracts, against a keyword-search baseline. The results determined the model every abstract in these databases is encoded with.

Brian E. Perron, Miao Wang, Nanyi Deng & Eunhye Ahn

Read the report