How Deterministic Pre-Filtering Cuts LLM Costs in Google Docs-to-CMS Pipelines

A software developer has detailed a technique for importing Google Docs into Contentful CMS without relying solely on large language models for content mapping. The approach introduces a deterministic pre-filtering stage that segments documents into regions based on headings and tab labels, then scores each region against available content types. Scoring is done using Jaccard similarity — a lightweight, synchronous set-intersection calculation — rather than embeddings or TF-IDF vectors, keeping the process fast and free of additional model calls. Each document region is ranked against content types and only the top candidates are passed to the LLM, significantly reducing context window usage and token costs. The method is designed to scale reliably even when a content model contains dozens of content types with many fields each.