D05 — A supplier's feed to a clean catalog

Trailhead Supply, an online outdoor-gear shop, lists the products of Alpenrand Outdoor, a German maker. Every day Alpenrand sends its product feed as a spreadsheet export — and every day a few rows are wrong: a category written in German, a price typed with a letter O instead of a zero, an article code left empty. This workflow publishes the good rows at once, and turns the bad ones into a list of corrections the supplier can act on, without anyone opening the file.

The reference run, on the feed of twelve products — 0.15 credits:

Delivered	What is in it
Shop catalog	7 products with SKU, title, category, retail price, stock, weight and a description
Corrections file	5 rows to fix, by line in the supplier’s file, with the request and the exact check errors
Email to the supplier	below

Subject: Corrections for tomorrow's product feed

Hello Alpenrand Outdoor,

Thank you for today's feed.

Rows received in today's feed: 12
Rows listed in the shop: 7

Please fix these rows and include the corrected versions in tomorrow's feed:

Line 5 (AR-3111): category: sent "Zelt", expected one of backpacks, tents, sleeping-bags, apparel, cookware
Line 7 (AR-4206): price_eur: sent "12,5O", expected a number
Line 9 (no article code): sku: sent "", expected AR- followed by four digits
Line 10 (AR-5310): stock: sent "-3", expected at least 0
Line 12 (AR-6121): price_eur: sent "", expected a number

Best,
Trailhead Supply — Purchasing

And one listing of the catalog:

SKU	Title	Price (EUR)	Description
AR-4205	Firnlicht -5 sleeping bag	263.95	The Firnlicht -5 sleeping bag is designed for nights down to a comfort rating of -5°C. Its responsibly sourced 650-fill down keeps weight to 0.98 kg, and the included compression sack packs it down for transport.

The case

The feed is a CSV exported from a German spreadsheet: semicolons between the columns and commas in the numbers.

sku;name;category;material;weight_kg;price_eur;stock;notes
AR-2101;Gratkamm 38 daypack;backpacks;recycled nylon 210D;0,86;42,50;64;hip belt removable, hydration sleeve
AR-3111;Hochmoor 1 tent;Zelt;silnylon, aluminium poles;1,48;129,00;8;solo version of Hochmoor 2
AR-4206;Firnlicht +5 sleeping bag;sleeping-bags;responsibly sourced down 650 fill;0,71;12,5O;22;comfort +5 C, summer weight

The shop’s contract for a row is short: an article code AR- and four digits, a name, one of five categories, a material, a weight between 0 and 25 kg, a wholesale price above zero and a whole, non-negative stock. The shop sells at 1.9 times the wholesale price, rounded to a price ending in .95.

Input	Who provides it	Example
Supplier feed	the supplier, every day	a CSV or Excel file uploaded as an asset

What it shows

• Read the whole row, then check it. Data Input reads the CSV with locale: de-DE, so 42,50 is the number 42.5 and 12,5O stays text. Data Rows could give each column its own typed port, but a single bad cell in a number column would then stop Data Rows itself (ROW_CONVERSION_FAILED): no row reaches the next steps, which are cancelled, and the run ends partial_success with nothing produced. Here each iteration passes its row — all its values as one JSON object — to Validate JSON Schema, and only that row is judged. See Spreadsheets and Schemas.
• The schema is the contract, written once. Codes, categories, numbers and limits are declared in the node’s inline schema, in route mode: a valid row continues on normalized_json, a rejected one is cut from the listing branch, and its valid and errors outputs drive the correction branch.
• Rules for money, AI for words. The retail price is round(wholesale * 1.9, 0) - 0.05 — 42.50 becomes 80.95 — and never passes through a model. The description is written by AI from the checked record only, told to use the material, weight and notes it is given and to invent nothing.
• Every problem goes back as something to do. For a rejected row, the AI reads the row as sent and the check’s errors and writes each problem in one fixed form — the column, the value sent, what is expected — so the supplier reads price_eur: sent "12,5O", expected a number instead of Value is "string" but should be "number". The exact errors stay next to it, in the corrections file, for anyone who wants them.
• Counts by rule, the email by AI. The rows to fix are collected, counted by Calculate, and the rows listed are the rows received minus those; the AI writes the email around these numbers and copies each request word for word. The line numbers are the lines of the supplier’s own file (row_index + 2: rows count from 0, after the header).

The workflow

The figure above tells the workflow as stages; the exact graph is the workflow (as a diagram), node by node in the table.

Node	Type	Why
Supplier feed	input/data, locale: de-DE	The day’s file; decimal commas read as numbers
Each row	enumerate/data_rows	One iteration per product; row carries all its values, sku and row_index are kept for the corrections
Check the row	utility/json_schema_validate, inline schema, route	The shop’s contract for a row
SKU, Name, Category, Wholesale price, Stock, Weight	utility/extract on normalized_json	The checked values, typed; present only for a valid row
Retail price	utility/calculate	round(wholesale * 1.9, 0) - 0.05
Row as text / Product copy	text/format json, ai/text_generation with low reasoning effort	The description, from the checked record
Shop catalog	aggregate/csv, whenAbsent: skip_row on SKU	One line per listed product; rejected rows leave no line
Only rejected rows	utility/filter, invert: true on valid	The errors of a rejected row, and nothing for a valid one
Row as sent / Errors as text / What to fix / Correction request	text/format json, text/template, ai/text_generation at temperature 0	One request per problem, in a fixed form
Line in the file	utility/calculate row_index + 2	The line the supplier sees in their spreadsheet
Corrections for the supplier	aggregate/csv, semicolons, whenAbsent: skip_row on the request	One line per rejected row
Rows to fix / Rows listed	aggregate/json, utility/calculate count(…) and received - to_fix	The numbers the email states
Email facts / Email to the supplier	text/format, text/template, ai/text_generation	The email, around the numbers and the requests
Outputs	output/csv ×2, output/text	The catalog, the corrections, the email

Reading a run: the catalog and the corrections always add up to the feed — here 7 and 5 of 12. In the trace, the listing steps of a rejected row and the correction steps of a valid row are skipped.

Cost: about a hundredth of a credit per AI text — a description per valid row, a request per rejected row, one email — so 0.15 credits for this feed of twelve. No description is written for a row that will not be listed.

Install it

The example installs into your workspace with an API key, through the public API only — see Installing an example:

node madoo-install-example.mjs d05-supplier-feed --api <your Madoo API URL> --key <key.json> --run

--run uploads the example feed and processes it (about 0.15 credits). To process your own, run the workflow with your file on the input Supplier feed.

Change it

• Your contract: the rules are the schema in Check the row — add a column, a pattern for EAN codes, a maximum price; a rejected row then explains the new rule on its own.
• Your pricing: the margin and the rounding are one expression in Retail price; a price per category is a Select by Key in front of it.
• Another supplier: set the locale of Supplier feed to the way the file writes numbers (it-IT, fr-FR, or none for dots); an Excel file needs no change.
• Images too: a text-to-image or product-photo step on the listing branch runs only for valid rows, like the description — see D01.
• Straight to the supplier: connect the email output to your mail system, or send the corrections file back through the supplier’s portal.

Files

manifest · workflow · the complete graph · the feed