Introduction

​

What is VectorLint?

VectorLint is a command-line tool that reviews and scores documentation using large language models (LLMs). Instead of regex patterns that can only catch surface-level issues, VectorLint uses an LLM-as-a-Judge approach to catch terminology misuse, technical inaccuracies, and style inconsistencies that require contextual understanding to detect. If you can write a prompt for it, you can lint it with VectorLint.

​

Why VectorLint exists

Traditional prose linters like Vale work by matching text against fixed regex patterns and word lists. They catch what you explicitly tell them to catch, but they can’t reason about meaning, context, or technical accuracy. VectorLint fills that gap. You define a rule once as a Markdown prompt, and the LLM applies it across your entire content library, scoring each document, surfacing specific violations, and explaining why each violation matters. This gives documentation teams something they haven’t had before: a shared, measurable definition of content quality that scales across writers, repositories, and content types.

​

What you can review

​

How scoring works

VectorLint scores each document using one of two methods, depending on how the rule is configured: Density scoring applies when a rule defines discrete issues to find (like a grammar checker). It calculates scores based on error density (issues per 100 words) so results are comparable across documents of any length. Rubric scoring applies when a rule defines quality criteria to rate on a spectrum (like tone or completeness). The LLM scores each criterion on a 1–4 scale, which VectorLint normalizes to a 1–10 scale for consistent reporting.

​

How false positives are reduced

VectorLint filters findings through a series of confidence checks before surfacing violations to help keep its output precise:

1. Candidate generation: the LLM returns all potential violations, each with structured evidence that VectorLint can verify.
2. Deterministic filtering: VectorLint applies confidence checks and only surfaces violations that meet all quality thresholds.

VectorLint’s output is intentionally stricter than what the model initially returns; it reports only findings that pass all confidence checks. You can tune how aggressively the pipeline filters findings to match your content workflow. See Handling false positives.

​

Next steps