Corpus
- USAU 2024 / WFDF 2025 rulebooks, segmented by section
- Per-section embeddings with version metadata
- Cross-rulebook section alignment for shared concepts
Group 08
· Capstone · Fall 2025
RAG Rules · Ultimate Frisbee — the rulebook, on the field.
A retrieval-augmented rules assistant for self-officiated Ultimate Frisbee. Players ask in plain language — "what's the call if the disc lands on the line?" — and the system answers by quoting the relevant rule, version-aware, citation-anchored, and explicit when the answer depends on a judgment call.
IWhat we built
What problem this solves.
Ultimate is one of the only competitive sports played without referees. Calls are made by the players themselves, which puts a real burden on knowing the rulebook well enough to dispute a call without slowing the game. The official text — USAU and WFDF — is long, occasionally contradictory across versions, and not designed to be consulted during a point.
Group 08 built for the captain holding their phone on the sideline. The product answers a question in two sentences, quotes the rule, and is explicit when the answer hinges on judgement that the rulebook itself defers to the players.
IIHow it works
The system, end to end.
The corpus is the canonical USAU and WFDF rulebooks, segmented by section and indexed by version. A query is retrieved against both corpora; the top passages are quoted verbatim in the answer with a citation back to (rulebook, section). A guard-rail layer detects judgement-call queries and is honest about the rulebook deferring to the players — instead of synthesising a confident answer that the rules themselves refuse to give.
The interface is built for the sideline. Answers are short, mobile-readable, and never longer than three sentences. The full rule text is one tap away when the captain wants to read the whole section, but the default is the answer to the question that was asked.
Ingest
Player question
plain-language ask
Transform
Retrieval
Pinecone, per-rulebook namespace
Storage
Rulebook corpus
USAU 2024 / WFDF 2025
Model
Answer contract
quote, don't paraphrase
Surface
Mobile UX
field-side, one-handed
Ingest Transform Model Storage Surface
IIIThe stack
What it's built on.
| Corpus | USAU 2024 / WFDF 2025 rulebooks, segmented by section |
|---|---|
| Retrieval | Pinecone vector index, separate namespaces per rulebook |
| Answer contract | Verbatim quote of the controlling sentence |
| Mobile UX | Sideline-readable typography, single-thumb layout |
Retrieval
- Pinecone vector index, separate namespaces per rulebook
- Per-query league selector — USAU / WFDF / either
- Top-3 sections surfaced with section IDs
Answer contract
- Verbatim quote of the controlling sentence
- ≤3-sentence summary above the quote
- Judgement-call detector triggers the deferral phrasing
Mobile UX
- Sideline-readable typography, single-thumb layout
- Offline cache of the rulebooks for connection-poor fields
- Tap-to-expand full-section view; back-to-game default
IVDeliverables
What the team shipped.
Source repository GitHub · code, tests, README
Demo video Capstone day · screen recording, 4–6 min
Write-up PDF Final brief · methods, evaluation, reflection
Slide deck Capstone presentation · 10 slides
VWhat sets it apart
What sets this capstone apart.
The rulebook in its own words.
Paraphrase is the failure mode in rules-RAG: the system invents a rule that almost-but-not-quite matches the source. The system quotes the section verbatim and links to the rest. The user reads the rule, not a model's version of it.
USAU and WFDF, separately.
The two rulebooks disagree on edge cases. The system asks which league before it answers — and surfaces both verdicts when the question crosses leagues. There is no "universal" answer where the rules themselves don't have one.
Some calls are not in the book.
Ultimate is self-officiated by design. When the rulebook defers to the players, the system says so — and quotes the deferral. The product never pretends to settle a call that the sport itself refuses to settle.
VIIInstructor note
How this project landed.
RAG over a rulebook is a project that fails the moment the model starts paraphrasing. Group 08 understood this from the proposal forward and committed to a verbatim-quote contract early. The first review reinforced it; the adversarial week did not break it.
The result is a small, useful, honest tool. It cites; it defers; it fits on a phone. The capstone teaches the lesson the rest of the cohort can take to any RAG project: when the source is canonical, paraphrase is the enemy.