Description parser

An offline LLM pipeline that reads creation descriptions and turns their load-order hints into rule books.

The problem

Creation authors routinely write things like "Load this after <X>" or "Requires <Y>" in their description. This information is invisible to LOOT and to the Categories sorter. Encoding it by hand into rule books would be a huge amount of manual work for ~5K creations.

The pipeline

Why offline and why LLM

The parser runs offline, in the project — not at runtime in the user's toolkit. Its output is a set of JSON rule books that ship with the release. End users never hit an LLM API.

Why LLM: description prose is full of noise, typos, creation names that look nothing like their plugin filenames, and relational phrasing ("install this first, but after the base mod") that naive regex approaches fail on. An LLM is the cheapest tool that handles this robustly.

Output confidence

The LLM is asked to emit a confidence level with each hint. High-confidence hints become rule book entries; low-confidence hints are logged for manual review and never ship as rules. This keeps the false-positive rate low — a bad rule would actively mis-sort users' load orders.

Relationship to rule books

The output is a curated rule book (see Rule books). It sits at the same priority as any other curated book, and is overridable by any user rule book. If the parser gets something wrong, users can fix it locally without waiting for a release.