AskEdith – Natural language interface for databases | Hacker News Reader

AskEdith – Natural language interface for databases

askedith.ai

85 pointsWallyFunk49 comments

Now I can stop tweaking my SQL query and start tweaking my prompt.

Less cynically but more critically, you’ll still have to review the SQL generated by this program. Unlike reviewing art, that review will be difficult, because reviewing SQL is not easy (unless the query was trivial to begin with).

Then you can buy the upgrade that will automatically create dummy data and write the tests for the query you just created ;-)

I'm looking forward to the day when there's an open source version of these models. I am not interested in making a gazillion API calls to a third-party vendor about how I'm introspecting into my own reports.

You have an unnecessary apostrophe after "CSV" on your Free tier pricing box. Apostrophes do not make initialisms plural.

TBH I disagree. In some fields it has become something of a convention, especially when the thing you're pluralizing is a variable or some other mathematical expression. For example I might write $x_1,x_2,...,x_n$, and then refer to the sequence as "the $x$'s". Happens all the time in professional journals by authors who know very well what the prescriptive rules of written English are. It is just easier to read this way.

> TBH I disagree.

You may, but it's incorrect. Try to tell o365 about the Oxford comma, I tried, it's futile. Best to assimilate and maintain pay-stream.

I've found I can sidestep that by including a label, such as "the $x$ items" "the `AWS_` variables"

In this specific case, I'd use "from your own CSV files" or "uploads" depending on which is more accurate

Its defiantly against the do's and donts of English, yes, apostrophe's seem to be the most mistaked parts of English. Also my favorite cigarette's are Marlboro 100's.

To say nothing of my personal landmine: plurals vs possessives of words that end in "s"

This is one of those products where it sounds cool but it will only work for very simple queries, where writing the prompt and reviewing the generated SQL query takes as much time as writing the SQL query from scratch.

Also, even the SQL shown in the EXAMPLES is bad :|

I guess it's true, you won't be writing SQL from scratch, you'll be writing SQL from very broken SQL.

Can't see the top sql because it keeps flipping over, WTF.

Where are these examples?

I copy-pasted them to be able to read them better.

    SELECT
        first_name,
        last_name,
        (
            SELECT MAX(order_total)
            FROM orders
            WHERE customer.id = orders.customer_id
        ) AS maximum_order
    FROM customers;

There's this one to get the maximum order amount for each customer that struck me, and another similar one.

Instead of doing a JOIN they do this weird subquery. You can clearly see that it's the work of an AI and how this AI works (putting together lego pieces which are subqueries), because nobody would be writing this query like that.

Plenty of people write SQL like that. It’s usually Python devs who are overly fascinated with shitty AI.

I’ve seen it a lot by older SQL devs or in enterprise places with ossified standard patterns. Lots of RDBMS’s historically had quirks and a lot of workarounds became cargo cult practices that often got culturally transmitted to the communities of other DBMSs and/or survived beyond the problem they were meant to workaround.

Am curious. How would you write it 'properly'? Other (good) alternative is

   with toporders as (
      select max(order_total) as maxOT, customer_id
      from orders
      group by customer_id
   )
   select last_name, maxOT
   from customers join toporders 
   on ...

What would you suggest?

Correlated subqueries run N times, once for each outer row. Your solution (assuming the database supports predicate pushdown) is far better.

  SELECT 
    customer.first_name, 
    customer.last_name, 
    max(order.order_total) as maximum_order
  FROM 
    customers customer
  INNER JOIN
    orders order
  ON (customer.id = order.customer_id)
  GROUP BY
    customer.id

I don't think subqueries are so bad - at least they're clear, and if fast, that's surely problem solved. IME they are clearer and often faster, so overall better. I know purists don't like them but I do.

In fairness there's nothing weird about that, it being just a typical correlated subquery. I'd quite possibly do it that way.

Right, weird wasn't the right word. It's bad (here, not always).

WHY bad? HOW am I supposed to learn anything from "It's bad"

Well, this isn't that bad on a real database which will know to decorrelate[1] the subquery but a GPT based AI probably won't advise you to do an EXPLAIN and check the plan.

1- e.g as Jamie explains here https://www.scattered-thoughts.net/writing/materialize-decor...

Now this is a useful answer, and perhaps the best in the thread, thanks

Also, it doesn't work.

I uploaded a very simple dataset in a CSV format, and not even the suggested questions after the upload works.

Demo is broken but it works once you login to a free acc

> Demo is broken

This is a deal breaker for me at least; if a service can't handle a predicted load from HN and friends, it's not trustworthy.

It's like buying a car: if I can't start it before a test-drive, I would never buy it.

I hear you, we're fixing it right now.

disclaimer: I worked on AskEdith

I don't think you understand.

As long as this is an SaaS, it's useless; no one should depend on a third-party provider to create VIEWs for them. "Oh, we can't deliver tax reports to the government because AskEdith is down" doesn't work as an excuse, at least not here in Norway.

It also prompts to make an account (and won't respond to the demo submit request) after what, 3 tries on the demo? This not really enough to determine whether it has the horsepower to do anything significant. Requiring signup to do even basic exploration of capability turns me away.

Echoing others' experiences, really difficult to try the demo. I saw data - I typed a question, hit the button. Nothing. I need to hit Enter. Do so. Wait for like a minute - progress bar appears? Then a message about datasets. Huh? ..

Awesome! I suspect that natural language will soon be the default way we interact with data, so I’m excited to see this. There’s just such a clear need for something like this and the technology is finally ready.

(Full disclosure: I have a side project exploring something similar. Demo: https://hermes-odmf.onrender.com. See the Metrics Catalog link for available metrics; all data is from 2018)

I'd really like to see studies comparing the accuracy of queries with such a system compared to manually written ones to know if I can actually trust it. Natural language is a great way to interact with data, but only if the intent is correctly translated. With AskEdith you at least get a SQL query that can potentially be reviewed for correctness.

Yes, this is a great point. To a first approximation, whether users trust the results is the only thing that matters for this kind of tech, imo.

Are there any DB or cloud vendors already offering natural-language to SQL capabilities? I'd be a bit surprised if Google or AWS doesn't have it, given their deep expertise in NLP.

[Shameless plug] If you wish to try another, I built a small proof-of-concept here: https://www.olym.pe

(registration necessary)

Not in a similar vein, but I find ultorg to be much better.

https://www.ultorg.com/

I'm trying out the demo with "What is the most typical experience level for people whose job is analyst?" and just get a half-way full red progress bar as a reply.

I tried it with "which region has the largest salary disparity?" and got the same red bar.

Demo is broken but it works once you login to a free acc

Wow very cool. Would love to see it interface with Bigquery or other data lakes so it could grab the schemas and structure query

I can see this being useful for managers or marketing people but not really for engineers.

I would really like to see how this would work against non trivial data (and queries)

The website's awful. There's no user testing on it - why do I have a 20 sec progress bar before it tells me I need to select files - where from? Drag from where? What format?

Found the datasets - what is AAPL.csv? What do the columns mean? How do I use them to query on? The UI has been designed by someone who doesn't know anything about UIs. Like, anything.

Added: "AskEdith translates English to SQL to help you save 1 hour every day". Well that's a bloody strong claim. Wonder if it's justified.

Where’s the anti-ORM crowd to defeat this idea?

This seems so far in the other direction that the same criticisms no longer apply. It would be like being against LCD screens in cars, and then complaining that Boeing's new 787 has flight displays.

I don't think I'm very anti-ORM, but I prefer using SQL where I feel for it. ORM is usually an overhead, and a bad implementation can make the ORM useless.

To answer your question: do we need to learn a QRM now? :) I think we are a few years away from that, ref. "what was the average salary for people earning more than X USD in 2019, adjusted for the purchasing power each person lives in?"

This is an example of a third-party data set that you don't want to be part of your core application, but maybe for a reporting system that talks to both your core application and some other third-party sources.

So. I'll stay with SQL until further notice, because a) this doesn't work, and b) third party data needs careful handling.

When you can only produce SQL via natural language prompts then how are you going to be able to troubleshoot your queries?

Silly question, but are you suggesting that in code you'd put:

sql_ai = "which customer spent the most with us in the past 12 months?";

Shh dont give them ideas. The PHP crowd will absolutely try to adopt it.

Reply on news.ycombinator.com