Hacker News Reader

Comment by helsinki | Hacker News Reader

I wonder why nobody uses jsonl format to represent an entire codebase? It’s what I do and LLMs seems to prefer it. In fact, an LLM suggested this strategy to me. Uses less characters, too.

View on HN

Are you suggesting that there's a correlation between what input formats provide best performance for an LLM input, and what sequence of tokens the same LLM outputs when prompted about what input formats provide best performance? Why would that be?

I don't think it's much difference, but I've read that Markdown codeblocks (or YAML, or XML) is better for code than JSON, for example: https://aider.chat/2024/08/14/code-in-json.html

I think it makes sense.

YAML is shorter and easier to read, Markdown codeblocks have no added syntax between the lines compared to normal code.

But JSON vs JSONL I can't come up with any big advantages for the LLM, it's mostly the same.

Why wouldn't that be? We've had several generations of LLMs since ChatGPT took the world by storm; current models are very much aware of LLMs that came before them, as well as associated discussions on how to best use them.

You can use JSON using the accept parameter of the API. The url structure remains the same. It also supports YAML and I found that's easiest to read for LLMs.

Previous example but in JSON:

https://uithub.com/upstash/docs/tree/main/qstash?accept=appl...

Is there any reason to prefer JSONL besides it being more efficient to edit? I'm happy to add it to my backlog if you think it has any advantages for LLMs

Reply on news.ycombinator.com