ingredient-parser

ingredient-parser is a Rust project for parsing recipe ingredient lines into structured data. This repository contains the core library and related tooling.

For full documentation, usage examples, and features, see the detailed ingredient-parser/readme.md.

The canonical backlog is GitHub Issues, tracked in the Ingredient parser Project. See the contributor workflow before starting work.

Maintainer toolkit

Run pnpm --filter @ingredient-parser/ui desktop:dev for the macOS Parser and Cookbooks desktop workspaces. See desktop setup and testing. The terminal interface has six command families: ingredient, amount, text, recipe, cookbook, and corpus. Use --help on a family or command to see its inputs. Results default to human-readable output, including when piped. Select --format json for structured output; ingredient batch also supports --format jsonl. Progress, errors, and parser traces go to stderr. corpus table writes HTML and does not support a structured format.

cargo run -p food-cli -- ingredient parse "1 cup flour, sifted"
cargo run -p food-cli -- ingredient parse "1 cup flour" --explain --format json
printf '1 cup flour\n2 tbsp sugar\n' | cargo run -p food-cli -- ingredient batch - --format jsonl
cargo run -p food-cli -- amount parse "2 cups" --format json
cargo run -p food-cli -- cookbook inspect book.epub
cargo run -p food-cli -- cookbook estimate book.epub
cargo run -p food-cli -- cookbook extract book.epub --out run.json
cargo run -p food-cli -- cookbook explain run.json --title "piecrust"

Ingredient batches retain every input line, including empty or review-needed lines, with a line number and an error field in structured output. A successful batch means all lines were processed; review flags remain visible per record.

The cookbook family reads an EPUB offline (inspect, explain), prices a run before spending anything (estimate), extracts through the AI gateway (extract), and re-runs a recorded dump with no credentials (replay). eval scores extractions against hand-authored answer keys, sample draws random recipes or books, and runs and models list saved runs and the model catalog. See cookbook extraction, or cargo run -p food-cli -- cookbook --help. Exit codes are 1 for operational errors, 2 for invalid invocations, 3 for incomplete extraction, and 4 for evaluation mismatches.

Generated EPUB test fixtures

cookbook-fixtures builds deterministic EPUB bytes in memory — cookbook_epub() plus fixtures covering real publisher layouts (split_spine, epub3_nav_pagebreaks, typographic_variations). They need no network and no extraction model. To write one out for manual inspection:

cargo run -p cookbook-fixtures --example write_fixture -- cookbook /tmp/synthetic-cookbook.epub

Frontend

ui/ is the one frontend: the public website (landing page plus the Parser and Cookbooks workbench) and the desktop window are the same build. Both run the same Rust command layer, food-core: natively in the desktop app, compiled to WebAssembly (food-wasm) in a worker on the web. The TypeScript contract ui/src/api/generated.ts is generated from Rust.

Run pnpm install --frozen-lockfile at the repository root, then:

• pnpm dev: build the WASM package and start the website at http://localhost:1420.
• pnpm --filter @ingredient-parser/ui desktop:dev: start the macOS app.
• pnpm build: build the WASM package and the frontend (ui/dist).
• pnpm lint / pnpm test: lint and unit-test the frontend.
• pnpm --filter @ingredient-parser/ui test:e2e: Playwright against the production build — the website in Chromium with real WASM, and the desktop shell in WebKit against a recorded fixture (see food-app).
• cargo run -p food-app --example export_bindings: regenerate the TypeScript contract.

The WASM build needs the wasm32-unknown-unknown target and wasm-pack.

The website is a Cloudflare Worker serving ui/dist as static assets (ui/wrangler.jsonc). CI deploys it to https://ingredient.nickysemenza.com on every push to main, and uploads a Worker Preview named pr-<number> for each pull request from this repository.