Engineering Notes

Notes on web and ML engineering from Greenskin Labs, including the Rules Oracle RAG series.

Series

Engineering the Rules Oracle

Modern tabletop games rely on massive, highly fragmented ecosystems of rulebooks, supplements, and constantly updating FAQs. When an obscure rule interaction or edge case arises mid game, it takes players out of the fun. I am building the Rules Oracle to solve this: a hosted Q&A engine that provides cited, page-referenced answers to complex rules questions. This series will cover my thoughts and learnings on engineering the initial work on the Rules Oracle.

Access is currently invite only during the beta period.

Part 1

In Good Character: Designing an Ingestion Pipeline for Hostile Tabletop Rules

June 10, 2026

Tabletop rulebooks are hostile to naive RAG: scanned PDFs, custom symbols, dense tables, and no reliable text layer. I used that mix as a validation corpus and built ingest for Structural Truth—heading paths, icon meanings in text, and book page numbers—by pivoting from unpdf and Document AI to Claude Haiku vision parse. Per-page and book-level caching makes re-chunking cheap; citation integrity belongs at the front of the pipeline.

RAGTypeScriptVision ParseClaudeData Engineering

Part 2

Pulling Rank: Using Fused Retrieval to Bridge the Alias Gap

June 12, 2026

Standard vector search has a vocabulary problem. It finds things that sound like your question. But when rules are written for a parent category and the question names a specific type, similarity is not enough. This second article in my series on the Rules Oracle details three parallel search lanes and the ingest-time supersession flag I use to bridge the Alias Gap. Finding the right rule is a matter of lanes and lexical safety nets, not just semantic proximity.

RAGTypeScriptHybrid RetrievalpgvectorRRFData Modeling

Part 3

Fielding Questions: Using JSON Constraints to Force Grounded Answers

June 12, 2026

At the table you need a ruling, not a confident guess. General-purpose AI fails that test on niche rulesets and will fake citations when you push back. This installment is about what I built instead: a system prompt that forces verbatim quotes before conclusions, so even a wrong answer still surfaces the rule text your group can read and decide from.

RAGTypeScriptPrompt EngineeringLLMAPI Design

Part 4

Preventing RAG Regressions: Eval Harnesses and Production Gates

June 15, 2026

The last installment splits in two. First: a cheap retrieval eval harness, query logging, and how I tell retrieval failures from generation failures without an automated generation suite. Second: the operational pieces that did not fit cleanly elsewhere, including provider swap, pre-flight gates, overrides, and the economics of a hosted side project.

RAGTypeScriptEvaluationObservabilityProduction

Structure Before Semantics: Local RAG for Adversarial PDFs

May 19, 2026

How constraint-aware design on a 6-year-old Intel Mac forced better retrieval architecture: two-pass tokenization, hierarchical chunking, graph-based RAG, and citation pruning — before any cloud infrastructure.

RAGTypeScriptembeddingslocal-LLMPDF-ingestionOllama