Robin Allenson
PRO
@robin
I built an AI for local SEO start-up in 2009, acquired 2014.
Now running Similar.ai & working on SiteTopic.io.
Yes, it's text-to-sql. I would just dump your entity relationships into the prompt and ask for the SQL for some example questions. Then, I'd take those and correct the SQL so that it works -- either asking the LLM to correct bugs you see or just doing it by hand. Then I'd put those input-output examples into the prompt and try it on new things and rinse & repeat.
o1-preview
will do best, but once you have a good prompt that works well, I would try that out on 4-o and smaller models
I like the page. Simple, clean. It was clear to me what the site did straight away.
I was expecting the dotted underlined text was a link to take me to an example.
Maybe you can find some way to make a few puns or use the cooking metaphor. e.g. you select the best ingredients, mix them up and cook up a new recipe? ...or maybe not 😜
With a lot of data, you can't stick everything in the prompt. You can make a selection of data first. It's very hip to use cosine similarity with vectors of your documents. It's great because it works with unstructured data. But if your data is structured or structure-able, you could just use SQL. That's nicer because it's less black box on what is going on and it's fast and cheap.
For example, take a text query and use an LLM with knowledge of your entity structure to write SQL, execute that, then use the results as context for your prompt. If you created a SQL db as an index of example (1), that would work. It would also work for example (2).
Hth!
Thanks so much! Very interesting. So that's sort of text-to-sql (or vice versa)?
Any resources where I can go down a bit of a rabbit hole for a few hours and learn more about it?
So my understanding would be that in case 2 I would send in the prompt itself something like
table: my_table
fields: id INT; name VARCHAR, email VARCHAR, country VARCHAR
And then if I e.g. ask how many people in my app are from the UK, I'd send the query + the table structure in the prompt and it would probably return something along the lines of SELECT count(id) FROM my_table WHERE country = UK, right?
So depending on the complexity of the database (and how often it changes), it still might be a bit tricky, especially if I have relational tables that connect multiple tables via IDs, etc. as this might need additional explanations in the prompt. If I'd use one of the big LLMs, such as OpenAI, what sort of model would work best with this?
Yes, it's text-to-sql. I would just dump your entity relationships into the prompt and ask for the SQL for some example questions. Then, I'd take those and correct the SQL so that it works -- either asking the LLM to correct bugs you see or just doing it by hand. Then I'd put those input-output examples into the prompt and try it on new things and rinse & repeat.
o1-preview
will do best, but once you have a good prompt that works well, I would try that out on 4-o and smaller models
We use local accountants who charge a flat fee and a fractional Head of Finance. He works a 2-3 days a month and handles all the enterprise invoicing, payroll, accounts payable stuff, and sets up finance systems like Revolut, Xero, Rippling and others. I think it's super valuable. I could imagine someone like that using this bundle as a channel to help lots of start-ups.
Thanks @alvivanco! I'm learning some of the WIP-language like builds and to-do's, which always seem to be "done". but it's nice to have a place where I can post the little things we shipped.
yes, the "todo's" are for things you do during the day for your projects.
If you wish to do a longer post, and share details on a specific project, you can go on your project's page and click on "Write a Milestone" (on the right side of the page)
I loved " I would focus on boosting the authority of the whole site and use proper internal linking". I've got lots of data to show that internal linking works. And naturally organically boosting the domain of the site coincidentally means that you are actually promoting your business too.
Cool! Happy to help. I'd love to hear how you get on.