Pretty neat initial launch. What's top of mind to add to it?

There is a PDF parser, LlamaParse, (which is open to everyone), and a managed ingestion/retrieval service, that is currently invite-only.

Planning broader releases in the future for sure.

The output is either text or markdown, and from there you can handle it however you need.

In LlamaIndex for example, there are a a few markdown-specific classes that work well with this.

You can find an example over in the repo -- https://github.com/run-llama/llama_parse/blob/main/examples/...

I was hoping to get structured data. For example, parsing an voice will give results like {"title"... "line_items": [...], "date":...}

This is actually what llama-index does if you don't use a vectordb integration

The problem is that LLMs aren’t very good at reasoning, needing constant guidance and double-checking.

Statistically likely answers are not necessary the correct ones.

Somebody else mentioned the phase of the moon when John Lennon died (which will be ruined in a few days when google indexs this site). There are results for that on the internet, google fails, Kagi fail, but Bings chat pulled the two different sources (one for the day and one for the phases of the moon).