A weekend ↓ architecture v1.0 · 2026-08-19

The allergen matrix that stays right: catch the reformulation nobody told you about

Your allergen chart is correct the day you write it - this reads every supplier sheet, works each dish out by arithmetic, and flags the ones that went stale

Setup: A weekend (~8-16 h) Running cost: $0-15/mo (software included) Time saved: Medium After setup: — Data lives: —

Document reading (OCR)Documents & archiveReminders

Built for small businesses in general — not a sector list.

Version history

v1.0 · 2026-08-19 — First release. Anchored on EU 1169/2011 Annex II (14 allergens, non-prepacked covered since 13/12/2014), UK PPDS labelling since 01/10/2021 plus the FSA written-first best-practice guidance of 05/03/2025 (recommended, not law), and California SB 68 / HSC 114093.5 in force 01/07/2026 (9 allergens, chains of 20+ locations only - Maryland HB 181 and Missouri HB 3446 would extend the same duty to every restaurant). Deliberate design line: software only ever ADDS an allergen, never removes one, and 'no evidence' is never 'free from' - the model is constrained to a closed list plus UNKNOWN and the rollup is set arithmetic.

An allergen chart is correct on the day you write it and slowly becomes fiction after that. This blueprint reads every supplier sheet once, works out each dish by arithmetic, and tells you which dishes went stale.

The architecture

the same flow as text

      [ Supplier spec sheets, labels, photos of packs ]

              ▼  (mailed to one address, or dropped in a folder)

      [ Docling — reads the page ]

              │  PDF/photo → text + table structure

              ▼

      [ Ollama, local model — maps wording to the allergen list ]

              │  "semolina" → cereals/gluten · "casein" → milk · unsure → UNKNOWN

              ▼

      [ Grist — the ingredient register ]  ←  one row per ingredient, evidence date, source file

              ▼  set union, in a formula column

      [ Recipes → dishes ]  a dish carries every allergen of every ingredient in it

              ├─ every ingredient verified and in date  → dish publishable

              ├─ any ingredient stale or UNKNOWN  → dish marked unverified

              └─ a new sheet contradicts the old one  → affected dishes listed

              ▼

      [ WeasyPrint ] → counter chart · PPDS labels · dated archive copy

      [ n8n ] → the review clock, the re-check, the flags to your phone

      Software adds allergens. Software never removes one. "No evidence" is not "free from".

That last line is the design, and I want to be blunt about why. If you ask a language model "does this dish contain milk?", it will answer, and it will sometimes answer no when the sheet was ambiguous. In an invoice check a false clean costs you money. Here a false clean is the whole failure. So the model gets one narrow job — reading a declaration and mapping wording to a fixed list — and it is allowed to say UNKNOWN. Everything downstream is set arithmetic in a spreadsheet formula. A dish is never marked clear because nothing was found; it is marked clear only when every ingredient in it has current, human-confirmed evidence.

The problem

The allergen matrix on your wall is a snapshot. Recipes get tweaked, a supplier is out of stock so someone grabs a different brand for a fortnight, and the manufacturer quietly reformulates a coating. None of those events walk over and update your chart. So the chart stays right for a while and then it is wrong, and nobody can tell you the date it stopped being right.

I am not going to quote you a reformulation rate. Nobody publishes one I would trust, and the mechanism does not need a rate to matter, because the cost of missing one change is not measured in percent. What is documented is the gap in the chain: the Codex Code of Practice on Food Allergen Management (CXC 80-2020) says manufacturers should have procedures for suppliers to notify them of formulation changes. Should. It is a recommendation, not a guarantee that an email lands in your inbox and gets read.

Meanwhile the obligation is real and it is on you. In the EU, Regulation 1169/2011 Annex II lists the 14 allergens and has required allergen information for non-prepacked food — restaurant meals, deli counters, bakery items — since 13 December 2014. In the UK, PPDS food packed on the premises where it is sold has needed a full ingredients list with allergens emphasised since 1 October 2021, and on 5 March 2025 the FSA published best-practice guidance pushing all non-prepacked food towards written-first allergen information. That guidance is strongly recommended, not law, and it is worth knowing the difference.

The US is arriving late and from the opposite end. There is still no federal rule for food prepared and served on site. California's SB 68, the Allergen Disclosure for Dining Experiences Act, added Health and Safety Code §114093.5 and took effect on 1 July 2026: written disclosure of the nine major allergens per menu item, on the menu or by QR code with a printed alternative. It only binds chains with 20 or more locations, so if you run one restaurant it does not apply to you. Read the bills copying it though — Maryland HB 181 and Missouri HB 3446, both introduced in 2026 — and they are drafted to cover every restaurant regardless of size. The phrase that keeps recurring in all of them is "knows or reasonably should know". That is a documentation standard, and documentation is the thing this blueprint builds.

Tool choices — and why

Docling

(MIT, 65k stars, pushed 18/08/2026) reads the documents. Spec sheets are tables, and a flat OCR dump of a table loses which declaration belongs to which product code. Docling was built to keep the table structure, runs locally, and handles photographed packs as well as PDFs.

Ollama

(MIT, 179k stars, pushed 18/08/2026) runs the local model that maps supplier wording onto the allergen list. This is the part worth having a model for, because declarations do not use the legal names — you get "semolina", "casein", "albumen", "tahini", "E322 (soya)". A 7–8B model on 16 GB of RAM handles it. Give it the 14 EU names and the 9 US names as a closed list, force it to return one of those or UNKNOWN, and never let it return a free-text answer.

Grist

(Apache-2.0, 11.5k stars, pushed 18/08/2026) is the register and the actual product here. Real formula columns are what make the rollup honest: a dish's allergen set is computed from its ingredients every time you look at it, not typed in once by a person who was in a hurry. NocoDB (64k stars) works too if that is what you already run.

WeasyPrint

(BSD-3-Clause, 9.5k stars, pushed 17/08/2026) turns the register into paper: the counter chart, the PPDS labels, and a dated PDF you keep. HTML and CSS in, PDF out, no browser. The dated copy matters more than the pretty one — if you are ever asked what your chart said in March, you want the March file.

n8n

(Sustainable Use License — free to self-host for your own business, 201k stars) is the plumbing and the clock. It watches the mailbox, calls Docling, writes to Grist, and runs the review cycle: flag any ingredient carrying an allergen whose evidence is older than six months, and anything else older than twelve.

Open Food Facts

(AGPL-3.0 server, ODbL data) is optional and it is a tripwire, not a source of truth. For barcoded products you can poll the API and watch allergens_tags and last_modified_t for a change. It is crowd-sourced, so a change there means "go look at the pack", never "update the register". Do not wire it to write anything.

Monthly cost

Software €0, all of it. On hardware you already own that is the whole bill. On a rented VPS with enough RAM for a local model, budget €10–15/month and check the current rate rather than trusting this sentence in six months. The commercial allergen and specification platforms in this space start well above that and are priced for manufacturers with a quality department.

Setup outline

1. List your top 20 ingredients by how many dishes they appear in. Start there. The long tail can wait a month.
2. Build the ingredient table first: supplier, product code, product name, allergens carried, evidence date, link to the source file. Evidence date is not optional — a register without it is the same wall chart with extra steps.
3. Collect the current spec sheets for those 20. Most suppliers will email them; for the rest, photograph the pack. This is the boring week and there is no way around it.
4. Wire Docling behind n8n and run those 20 through. Read every output yourself and correct it. You are building your prompt against real sheets, and you will find at least one supplier whose PDF is a scan of a fax.
5. Add recipes and dishes, then write the rollup as a formula: dish allergens = union of ingredient allergens. If a sub-recipe goes into three dishes, model it as an ingredient so the union chains properly.
6. Add the two status columns before you print anything: verified (a human confirmed this ingredient row against a sheet, on this date) and publishable (every ingredient in this dish is verified and in date). Only publishable dishes go on a printed chart.
7. Generate the chart and the labels with WeasyPrint, and archive a dated PDF on every regeneration.
8. Turn on the review clock: six months for allergen-carrying ingredients, twelve for everything else, plus an immediate flag on any new sheet, any supplier switch, and any delivery whose packaging looks different.
9. Handle cross-contact separately. That is a kitchen question — shared fryer, shared slicer, shared prep surface — and it belongs in a column you fill in by walking the kitchen, not in anything a document reader produces.

Pitfalls — the real ones

The substitution nobody logs

Supplier is out, the chef buys a different brand at the cash-and-carry, the dish is served for two weeks and the register never hears about it. This is the single most likely way your matrix goes wrong, and it is a purchasing habit, not a software problem. If deliveries come in through the goods-in check, tie a new product code to a new ingredient row and let the dish drop to unverified until someone looks.

Do not let the model say "free from"

Constrain it to a closed list plus UNKNOWN, and treat UNKNOWN as blocking. The temptation to accept a confident-sounding negative is exactly the temptation to avoid.

"May contain" is not the same as "contains", and it is not regulated the same way either

Keep precautionary statements in their own column. Rolling them into the allergen set makes every dish look dangerous and staff stop reading the chart, which is worse than not having one.

Sub-recipes are where the union breaks

Your house dressing goes into four salads. If it is not modelled as an ingredient with its own rollup, the mustard in it silently disappears from all four.

A chart is not a conversation

Under the FSA's March 2025 guidance the written information supports a trained member of staff, it does not replace one. The same is true wherever you trade. This build makes the written half correct and current; the training is still yours to do.

Keep the source file, not just the extracted flags

The PDF or photo the declaration came from is your evidence. Storage is cheap and a row saying "milk, checked 12/03" with no document behind it proves nothing.

This is not legal advice and it does not make you compliant

Requirements differ by country and by how you sell — prepacked, PPDS, loose, distance selling — and they are moving in the US right now. Check yours with your own food-safety authority. What this gives you is an accurate, dated, defensible record to comply with.

Verified repos

Docling — MIT, 65k stars, active (August 2026)
Ollama — MIT, 179k stars, active (August 2026)
Grist — Apache-2.0, 11.5k stars, active (August 2026)
WeasyPrint — BSD-3-Clause, 9.5k stars, active (August 2026)
n8n — Sustainable Use License (free self-hosted internal use), 201k stars, active
Open Food Facts — AGPL-3.0 server, ODbL data, active (August 2026) — optional tripwire only
NocoDB — 64k stars, active (August 2026) — alternative to Grist

Get the next blueprint as it publishes — free, practical, verified tooling only.

Subscribe · Want it audited or extended? DM @RXed_EU