Academic Metrics

The package collects publication records through Crossref, enriches incomplete metadata, deduplicates articles, and coordinates classification and export through one pipeline. The goal is reusable structured data rather than a one-off dashboard.

The classifier does not ask for an unconstrained topic label. It traverses the NSF NCSES hierarchy one level at a time, carrying the selected parent into the next decision so every leaf has a valid path.

Before classification, separate stages extract methods, analyze sentence structure, and summarize the abstract. Those outputs give the classifier a smaller and more explicit evidence surface.

When the model returns a label outside the allowed children, the retry path tokenizes the rejected output and applies logit bias against those tokens. The system actively prevents the same invalid answer from recurring instead of dropping the record or normalizing it after the fact.

Academic Metrics ships as an installable Python package with a CLI, storage adapters, export paths, and full Sphinx documentation. The web demo is a separate collaborator-owned consumer; the reusable package is the project presented here.

The LLM stages run through an earlier embedded version of ChainComposer. Academic Metrics was its production proving ground: the classifier declares method extraction, sentence analysis, summarization, recursive classification, and theme recognition across three managers, then carries named outputs into later prompts. That implementation was later extracted and published as its own package.