Show HN: 30k IKEA items in flat text

https://huggingface.co/datasets/tsazan/ikea-us-commercetxt

55•tsazan•1mo ago

OP here.

I took the unofficial IKEA US dataset (originally scraped by jeffreyszhou) and converted all 30,511 products into a flat, markdown-like protocol called CommerceTXT.

The goal: See if a flatter structure is more efficient for LLM context windows.

The results: - Size: 30k products across 632 categories. - Efficiency: The text version uses ~24% fewer tokens (3.6M saved total) compared to the equivalent minified JSON. - Structure: Files are organized in folders (e.g. /products/category/), which helps with testing hierarchical retrieval routers.

The link goes to the dataset on Hugging Face which has the full benchmarks.

Parser code is here: https://github.com/commercetxt/commercetxt

Happy to answer questions about the conversion logic!

Comments

colinbartlett•3w ago

Any practical use for this IKEA data specifically?

Or just a handy open data set you could use to prove out the concept?

DennisP•3w ago

I assumed it's because IKEA is famous for flat packing its furniture.

tsazan•3w ago

Exactly! IKEA removes the air from the box to save space, CommerceTXT removes the HTML/JSON bloat to save tokens. You made my day!

embedding-shape•3w ago

> IKEA removes the air from the box to save space

Huh? I don't think that's true, there usually is some sort of structural elements inside of the package, meant to be thrown away (usually made with cardboard/paper), and all Ikea boxes definitively have lots of air inside of them, not sure what would make you say otherwise, unless it's some joke I'm missing?

jayknight•3w ago

A box that contained a fully assembled kitchen table would contain a lot more air. I think that comment just meant IKEA designs items that can be packaged into a minimal volume.

embedding-shape•3w ago

Ah yes, on second reading it's actually pretty obvious that is what parent meant and I was reading it too literally. Thanks for the clarification, that's certainly correct :)

WildGreenLeave•3w ago

I've had the idea to setup an AI that automatically (re)designs a room using IKEA stuff. It would definitely help me decorate my room in a better way.

tsazan•3w ago

That`s great use case. If you ship it, let me know!

vachina•3w ago

There’s already a schema.org spec that defines a JSON-LD structured data that you can embed on every of your product page to provide a machine readable interface of your product.

For example, Google’s indexers already use this to surface pricing data. https://developers.google.com/search/docs/appearance/structu...

tsazan•3w ago

That`s is valid for search engines. But if JSON-LD was sufficient for agents, Google wouldn't have launched UCP (Universal Commerce Protocol) yesterday.

vachina•3w ago

Took a look, UCP looks like presenting an entire shopping lifecycle for agents.

JSON-LD is just read-only metadata for machines.

tsazan•3w ago

True. But extracting that metadata requires parsing the full DOM. CommerceTXT is for efficient discovery. Scan inventory cheaply first, then commit to the transaction.

sognetic•3w ago

Interesting! So did you do any experiments on a relevant subset of the data to test whether LLM performance degrades by introducing a new, presumably unknown to the LLM, format?

tsazan•3w ago

The 24% token savings come from converting JSON syntax to CommerceTXT.

reddalo•3w ago

I don't understand why new proposed standards are still polluting the root namespace (also see llms.txt).

These things should be put under /.well-known [1], not in the root.

[1] https://en.wikipedia.org/wiki/Well-known_URI

dkdcio•3w ago

I was not aware you shouldn’t do that — what’s the rationale/historical context?

embedding-shape•3w ago

Like most standards: "Because it's a standard". Kind of like setting a .body for a GET request, you can kind of do that, but why not do it the way it's intended to instead? Use POST :)

gunalx•3w ago

I have seen post being used instead of get, because of having encrypted parameters by default.

embedding-shape•3w ago

Yeah, and also because of firewalls sometimes stripping body of GET requests (not responses mind you, we're talking requests) to a server, and also because it's really uncommon to put a body on a GET request ;)

ljm•3w ago

Sending a URL encoded form or some JSON in a POST request is also easier for most people to understand than the myriad ways you might format a query string in the URL (which may have a stricter limit on size).

You only have to look at how different services handle arrays in query strings to understand that serialising it is conceptually easier.

Comes up a lot in search or filter APIs. I'm sure there was some effort many moons ago to create a QUERY method for that.

buildbuildbuild•3w ago

User friendliness. I’ve seen several less-technical people able to quickly access, create, and understand “llms.txt”.

It’s not ideal but representative of the tension between user experience and technical correctness.

reddalo•3w ago

>less-technical people able to quickly access

Why would somebody even want to access that file? It doesn't make any sense to make that more user friendly, it's for LLMs.

JosephRedfern•3w ago

I've heard that LLMs can perform worse with these more efficient representations compared to e.g. JSON, because they've seen far fewer examples of them during training. Do you know how true that is?

TechSquidTV•3w ago

Absolutely, but usually when working with a bespoke format for optimization, it's paired with an LLM specifically trained on that format.

tsazan•3w ago

You are right about cryptic formats. CommerceTXT is semantically structured. Models like GPT, Claude and Gemini understand it out-of-the-box via ICL.

btrettel•3w ago

Interesting. I had been thinking recently about grep-friendly structured text file formats given the constraints of regex. But I hadn't considered that you could design a structured text file format to be LLM-friendly given token constraints.

tsazan•3w ago

You're right.If a format is easy to grep, it is almost always cheap to tokenize. We treat token density as a primary design constraint.

usefulposter•3w ago

"OP here" is the funniest tell that shows up when using an LLM to write a post for HN or Reddit.

It's funny because it makes zero sense in the body of an initial post!

In comments replying to people downthread - maybe. But opening a top-level post with "Original Poster here" is just silly and shows a lack of respect for community etiquette.

https://hn.algolia.com/?dateRange=pastYear&page=0&prefix=tru...

dkoy•3w ago

Good catch, think you’re on to something

tokai•3w ago

I just understand it as lightly humorous. Like starting a anecdote with

>be me

Seeing it as a lack of respect is a huge stretch. And kinda conceited that you accuse someone of such, on the basis of a two word opener.

croisillon•3w ago

years ago i did a small tool that, when you entered a product number, would scan all IKEA-websites with currency Euro and return the prices for each of them ; not that i expected furniture tourism to become a thing but it was funny

tsazan•3w ago

Reminds me of a friend who built a comment sentiment analyzer years ago. At the time, it looked like great innovation...

bleonard•3w ago

A blast from the past. When Taskrabbit was acquired by IKEA, I built several tools that went through the whole catalog via various crawling approaches. One tool was to estimate how long it would be to put each item together for an initial training set.

chuckadams•3w ago

Well of course it would be flat text... ;)

I Write Games in C (yes, C)

We Mourn Our Craft

SectorC: A C Compiler in 512 bytes

Brookhaven Lab's RHIC Concludes 25-Year Run with Final Collisions

Hoot: Scheme on WebAssembly

The AI boom is causing shortages everywhere else

Stories from 25 Years of Software Development

OpenCiv3: Open-source, cross-platform reimagining of Civilization III

U.S. Jobs Disappear at Fastest January Pace Since Great Recession

Al Lowe on model trains, funny deaths and working with Disney

The Waymo World Model

Reinforcement Learning from Human Feedback

Start all of your commands with a comma (2009)

Vocal Guide – belt sing without killing yourself

Show HN: I saw this cool navigation reveal, so I made a simple HTML+CSS version

Coding agents have replaced every framework I used

France's homegrown open source online office suite

Selection Rather Than Prediction

A Fresh Look at IBM 3270 Information Display System

72M Points of Interest

History and Timeline of the Proco Rat Pedal (2021)

Unseen Footage of Atari Battlezone Arcade Cabinet Production

Where did all the starships go?

Show HN: Look Ma, No Linux: Shell, App Installer, Vi, Cc on ESP32-S3 / BreezyBox

Learning from context is harder than we thought

Monty: A minimal, secure Python interpreter written in Rust for use by AI

Hackers (1995) Animated Experience

Making geo joins faster with H3 indexes

Sheldon Brown's Bicycle Technical Info

Show HN: Kappal – CLI to Run Docker Compose YML on Kubernetes for Local Dev

I Write Games in C (yes, C)

We Mourn Our Craft

SectorC: A C Compiler in 512 bytes

Brookhaven Lab's RHIC Concludes 25-Year Run with Final Collisions

Hoot: Scheme on WebAssembly

The AI boom is causing shortages everywhere else

Stories from 25 Years of Software Development

OpenCiv3: Open-source, cross-platform reimagining of Civilization III

U.S. Jobs Disappear at Fastest January Pace Since Great Recession

Al Lowe on model trains, funny deaths and working with Disney

The Waymo World Model

Reinforcement Learning from Human Feedback

Start all of your commands with a comma (2009)

Vocal Guide – belt sing without killing yourself

Show HN: I saw this cool navigation reveal, so I made a simple HTML+CSS version

Coding agents have replaced every framework I used

France's homegrown open source online office suite

Selection Rather Than Prediction

A Fresh Look at IBM 3270 Information Display System

72M Points of Interest

History and Timeline of the Proco Rat Pedal (2021)

Unseen Footage of Atari Battlezone Arcade Cabinet Production

Where did all the starships go?

Show HN: Look Ma, No Linux: Shell, App Installer, Vi, Cc on ESP32-S3 / BreezyBox

Learning from context is harder than we thought

Monty: A minimal, secure Python interpreter written in Rust for use by AI

Hackers (1995) Animated Experience

Making geo joins faster with H3 indexes

Sheldon Brown's Bicycle Technical Info

Show HN: Kappal – CLI to Run Docker Compose YML on Kubernetes for Local Dev

Show HN: 30k IKEA items in flat text

Comments