Showcase

What people
actually built.

Not case studies. Nobody paid us and nobody approved the copy. Each entry says what broke first, because that's the part that saves you a weekend.

4.1M

Documents in the largest known index

61

Projects listed in SHOWCASE.md

30

Community connectors and parsers

0

Entries we were paid to include

Public records · @archive-collective

Four million scanned pages, searchable offline.

A volunteer group digitising municipal records from 1948 onward. Everything is a scan, much of it typewritten, some of it handwritten in margins. They run Quire fully air-gapped on a donated workstation.

What broke first: OCR confidence. Low-quality scans produced spans that read as plausible English but were wrong at the word level, and citations pointed at them confidently. They now index OCR confidence per span and refuse to cite anything under 0.7.

offline = trueOCR alwaysSQLite70B local

"We needed something that would still work in ten years with no company behind it. That ruled out everything except a file format and a license."

ACMaintainer, Archive Collective
Contributed the OCR confidence patch

"The refusal behaviour is the only reason legal signed off. It can decline, and declining is logged."

HMHalden ops team
Contributed the CI recipe
Docs QA · @halden-ops

A build that fails when the handbook stops being true.

Sixty questions their internal handbook is supposed to answer, checked on every merge with --strict. A page that quietly went stale now breaks CI instead of a new hire's first week.

What broke first: flakiness. A local model gave slightly different answers run to run, and the check failed randomly. Fixing it meant pinning temperature to zero and asserting on grounding instead of answer text. Assert that it can answer, not what it answers.

GitHub Actionsstricttemp 0
Science · @orrery-lab

Ten years of lab notebooks, asked in plain language.

A materials group indexed a decade of notebooks, instrument logs and rejected drafts. The valuable queries turned out to be about failures: which synthesis routes were tried and abandoned, and why.

What broke first: handwriting, and then units. The parser read "0.5 mL" and "0.5 ml" and "½ mL" as different things, which made numeric comparisons useless until someone contributed a unit normalizer. That's now in core.

custom parserunit normalizerpgvector

"It found the route we abandoned in 2019 and the reason, which was in a footnote nobody had opened since."

OLPostdoc, Orrery Lab
Contributed the unit normalizer
Also in the list

Fifty-eight more.

Municipal meeting search

Council minutes and agendas for a city of 90,000, with a public web UI behind a read-only proxy.

serve

Personal wiki

Ten years of Obsidian notes. The most-used deployment shape by a wide margin: one person, one folder.

local

Fan wiki archive

A 40,000-page wiki that went read-only. Preserved and made answerable, with citations to the last live revision.

html parser

Open-source triage

Answers "has this been reported?" against ten years of issues before a maintainer reads the new one.

agents

Course materials

A university department serving cited answers to students, with refusal turned all the way up.

strict

Incident review

Post-mortems, alerts and chat logs indexed together so "have we seen this?" is a query, not a memory test.

connectors

The full list is SHOWCASE.md in the repo. Adding yourself is a pull request, and we merge them without asking what you do for a living.

The ones that didn't work

Two shapes fail reliably.

Corpora that are too thin. A dozen pages of notes gives retrieval nothing to work with, and the result is a chatbot with extra steps. If you can read your whole corpus in an afternoon, you don't need this.

Real-time data. People try to point it at a metrics dashboard or a live database and get frustrated. Quire answers from documents, and a document that changes every second isn't one. Use a query tool.

Both are in the FAQ now because both showed up repeatedly in Discord before anyone wrote them down.

Add what you built.

A paragraph, your config, and what broke first. No approval process, no logo requirements, no case-study interview.