Industries

Resource Hub

How Text to SQL can replace your company analysts

By

Aleks Tiupikov

Jan 31, 2024

Skip to:

Share

What's Text to SQL? Is it still a thing? No, you can't replace nobody, the tech is not there!

All that and many more we've been told for the last year. Yet our opinions remain very polarized.

In this article, I'll show you why we believe that you can and probably should prioritize adopting AI in your organization instead of hiring more analysts.

We spent 12 months building a text to sql tool that can do 70% of the work most analysts do nowadays. And not only us, hundreds of others work 100-hour weeks to push this bar further and further.

This is happening, whether you like it or not.

Today you'll learn:

  • What Text to SQL is all about

  • How LLMs completely flipped the text-to-SQL movement

  • What is going to happen to the BI industry

  • What current solutions can exactly do

And so much more! Let's dive in… but first

TLDR - You Can’t Replace Them In One Night

I know you might be here thinking, "yeah, obviously you can't," but just hang on for a second. The idea here (which took us a year to understand) is that you CAN replace them, it's just not instantaneous.

But we tried. I mean, we really did think that when ChatGPT came out, it was the day. Not even close.

If you simply take an OpenAI API key, as many might think, and connect it to the front-end chatbot and try asking it some analytical questions (also providing your database information), in 50% of the cases, it will not work.

Why? Because:

  1. Your database is too big and can't fit in the prompt.

  2. The user’s query is too complex, and AI can't fit it all in one query.

  3. You're using GPT-3.5, which simply doesn't work.

  4. Your database structure is so messed up that AI has no idea how to translate it into relevant code.

  5. You don't understand how AI works, so the way you phrase your questions will make them never be answered.

  6. Hope you get the point.

Some of the fixes to the above require just extra coding. Some require extra training. But some require extra something that we had no idea about.

But in the end, if you connect it all together... it might work. And we see it's working pretty darn well!

What Is Text to SQL And How Does It Work?

Text to SQL is a natural language processing technology that converts plain text queries into structured SQL queries. It works by analyzing the text, identifying key elements such as entities and relationships, and transforming them into SQL.

The idea behind it is bread and butter - how to make people who do not know how to write code achieve things people who typically know how to write code do achieve.

And no, it's not new by any means. The BI (Business Intelligence) trend started in the 1950s with exactly the same idea.

The issue is it hasn't moved anywhere from there for the last 70 years. Obviously, there were some improvements and breakthroughs, but the core concept remained the same - translating business requests into SQL code using a user interface.

The whole BI trend was progressing, not because it was really innovating, but because the tech behind it was becoming better and better. Computers became faster, processing time shorter, interface cleaner.

But in a nutshell, it remained the same interface translation into generated sql queries. And nobody could do anything about it as it was a reliable and honestly single reliable and scalable way to receive insights from the dataset in a low latency way.

And as you know, different strokes for different folks, so it didn’t take too long for the alternatives to come up.

Text to SQL handles that by using LLMs to translate user requests into SQL code, so you don’t have to spend hours configuring the dashboard in Tableau.

It’s as simple as asking a question in english and getting an answer… or not?

How LLMs completely flipped the text-to-sql movement

When OpenAI released GPT-3.5, it changed the whole scene for everyone working on the text-to-SQL technology.

Not only did the overall accuracy of queries become better as the model had an immense knowledge difference compared to its predecessors, but also the semantic understanding of the natural language questions was raised significantly.

But GPT-4 was the one, in my opinion, that made all the difference. And the difference is in the decision-making process it can offer.

Now let me explain what I mean here. When I say decision-making in this context, I mean the actual decisions the model is capable of making to resolve your request in the best way possible (at least for the model).

In simple terms, what that means is the model can decide how to approach your question. A lot of questions are typically solved in a couple of steps. Now, GPT-4 can do it.

How GPT-4 decides to approach your question

Let’s say you’re analyzing the behavior of customers that recently signed up for your product. There're always going to be people that behave… very strange.

They can purchase a subscription in 1 second or they can be inactive for one month, then come back and get an annual membership.

The key here would be consistency. Is there a pattern in people doing that? If not - you shouldn’t factor them in your end analysis; otherwise, the end results will be skewed.

In statistical terms, these people are called outliers. And to determine who are outliers and who are not, you’d typically go through a couple of steps (I am using z-score method in this example):

  1. Calculating the historical mean

  2. Calculating the Standard Deviation

  3. Computing Z-score and pinpointing the customers beyond the threshold

This is a very common thing analysts would typically do while providing you with some report.

Now if you want to do this using Tableau, it would be really hard, since you will have to create a lot of custom properties.

And before GPT-4, any existing Text to SQL models, unless trained on this specific kind of question, would really struggle to generate it from scratch, since the query is insanely huge.

But GPT-4 completely changed the game in this regard. Now it can outline for you what should be done to spot the outliers and go do it step by step. As if it was a human analyst doing the work.

And since the queries, still being big, are logically not very complex - you will be able to get the answer in probably 2 minutes.

What is going to happen to the BI industry

Those of you working closely with BI tech might be wondering: “So you’re saying PowerBI, Tableau, and others will fade away while AI will conquer the world?”.

Practically, yes, you will not find these tools in the way they work right now in 5-7 years. Will they be replaced? Partially.

Mostly because these companies are not stupid and are already starting to adopt Text to SQL to slowly move away from the old user interface architecture.

Why did I say partially then? Well, that’s because since they have been optimized and developed with the old architecture in mind for the last 20 years, it would be really hard to adapt it to the whole new reality of words.

So most likely they’re just going to plug the LLM model on top of their stack and train it to operate with it. Which will probably work, but to the extent of how limited their existing stack is.

I doubt if the above example with outliers will ever be easily solved by any major BI tools without custom pre-configurations.

So in my opinion, there’s going to be a new wave (which is already happening) of Text-to-SQL analytics alternatives that will try to push out the big fish in the market. Whereas companies like Microsoft, Google will be trying to integrate AI in their analytics software which is going to take a long time.

The question is if any of them will be able to do it in the way that it does actually work?

So what can Text to SQL actually do right now?

Providing Answers Out of the Box

Right now, models can give you pretty good quality answers out of the box while feeding them the information about your company and your database correctly.

It does still require some minimal mapping, but it’s even hard to call it mapping. All you should do is to explain to AI the columns in your database that are unique to your business (like abbreviations) so then when you use these words, the model can “map” them to the actual fields.

This allows the models to provide answers out of the box without the need for extensive documentation or hard-coded definitions.

Multi-Steps Execution

As we talked about earlier, the advanced models can now generate a strategy for the upcoming steps they will need to undertake to get the data you’re asking for.

This also engages the user in the loop, so then it’s easy to follow and understand what AI is trying to do, adjusting it on the way.

Surprisingly, but as we discovered, the majority of mistakes AI do when trying to generate you a query is in misunderstanding your initial request. So if you can spot and correct it, the results will be really accurate.

Multi-steps execution of the requests facilitates exactly that.

Generating Charts

It goes without saying that if it can find data for you, it can present it in any way you can imagine. GPT-4 data analysis module is really good at plotting different kinds of charts, so you realistically will not need to do it in Excel anymore.

If you ask me how to deal with the charts that you want to keep on a dashboard and update every day, you can ask the model to generate you a chart based on the parameters provided and then simply save these parameters as well as a query for future users.

Semantic Layer

What if you say you already have so many queries created in different software that it would be nuts to redo it in a new tool even with AI helping you along the way.

That concern is tackled by the thing called a semantic layer, or what you can call a knowledge vault of your model.

Basically, you can import all of the queries you have in your SQL clients into it alongside with definitions of what each query means for the model to then use it in the future requests.

Let’s say you have a custom query and formula for lifetime value calculation of your customers. You can achieve the same by guiding the AI to it every time which becomes pretty annoying at some point.

Or you can define it in your semantic layer and then every time you’d ask “What’s the LTV of Q1 cohort?” kind of question the model will not try to generate anything from scratch but rather use your example to only adjust it based on the conditions in your request (e.g., timeframe or segmentation).

How Do You Know All That and Where Can We Try It?

We know that because we’ve been working on this the whole of 2024.

We tried so many things to see what works and what doesn’t to finally come up with a version that can actually do all the things described above. And this is probably the only product that can do it for now 🤞

A week ago, we released a free tool for everyone to use that can turn questions into queries only by knowing your schema.

We don’t charge for it, don’t even ask for an email. Only because we want to show you that it actually does work. You can try it here: datalynx.ai/text-to-sql.

If you want to try the product that can do everything mentioned above, you can sign up for the waitlist, which already has the trust of over 1000 people: Join the waitlist.

And if you have any suggestions or would like to help us move this project further, please do feel free to reach out at aleks@datalynx.ai.

I hope this article made you look at the problem of no-code analytics from a new perspective and raised a curious spirit to go give Text to SQL solutions a shot.

And even if it’s not our tool, I’m really pumped to see you joining this whole motion.

You rock! 😎

Here's some other great resources on this topic

  1. BirdSQL - Model to Human Level comparison benchmark

  2. SQLCoder-70b - New model that beats GPT4 based on CodeLlama-70B

  3. Spider - Another great benchmark list

What's Text to SQL? Is it still a thing? No, you can't replace nobody, the tech is not there!

All that and many more we've been told for the last year. Yet our opinions remain very polarized.

In this article, I'll show you why we believe that you can and probably should prioritize adopting AI in your organization instead of hiring more analysts.

We spent 12 months building a text to sql tool that can do 70% of the work most analysts do nowadays. And not only us, hundreds of others work 100-hour weeks to push this bar further and further.

This is happening, whether you like it or not.

Today you'll learn:

  • What Text to SQL is all about

  • How LLMs completely flipped the text-to-SQL movement

  • What is going to happen to the BI industry

  • What current solutions can exactly do

And so much more! Let's dive in… but first

TLDR - You Can’t Replace Them In One Night

I know you might be here thinking, "yeah, obviously you can't," but just hang on for a second. The idea here (which took us a year to understand) is that you CAN replace them, it's just not instantaneous.

But we tried. I mean, we really did think that when ChatGPT came out, it was the day. Not even close.

If you simply take an OpenAI API key, as many might think, and connect it to the front-end chatbot and try asking it some analytical questions (also providing your database information), in 50% of the cases, it will not work.

Why? Because:

  1. Your database is too big and can't fit in the prompt.

  2. The user’s query is too complex, and AI can't fit it all in one query.

  3. You're using GPT-3.5, which simply doesn't work.

  4. Your database structure is so messed up that AI has no idea how to translate it into relevant code.

  5. You don't understand how AI works, so the way you phrase your questions will make them never be answered.

  6. Hope you get the point.

Some of the fixes to the above require just extra coding. Some require extra training. But some require extra something that we had no idea about.

But in the end, if you connect it all together... it might work. And we see it's working pretty darn well!

What Is Text to SQL And How Does It Work?

Text to SQL is a natural language processing technology that converts plain text queries into structured SQL queries. It works by analyzing the text, identifying key elements such as entities and relationships, and transforming them into SQL.

The idea behind it is bread and butter - how to make people who do not know how to write code achieve things people who typically know how to write code do achieve.

And no, it's not new by any means. The BI (Business Intelligence) trend started in the 1950s with exactly the same idea.

The issue is it hasn't moved anywhere from there for the last 70 years. Obviously, there were some improvements and breakthroughs, but the core concept remained the same - translating business requests into SQL code using a user interface.

The whole BI trend was progressing, not because it was really innovating, but because the tech behind it was becoming better and better. Computers became faster, processing time shorter, interface cleaner.

But in a nutshell, it remained the same interface translation into generated sql queries. And nobody could do anything about it as it was a reliable and honestly single reliable and scalable way to receive insights from the dataset in a low latency way.

And as you know, different strokes for different folks, so it didn’t take too long for the alternatives to come up.

Text to SQL handles that by using LLMs to translate user requests into SQL code, so you don’t have to spend hours configuring the dashboard in Tableau.

It’s as simple as asking a question in english and getting an answer… or not?

How LLMs completely flipped the text-to-sql movement

When OpenAI released GPT-3.5, it changed the whole scene for everyone working on the text-to-SQL technology.

Not only did the overall accuracy of queries become better as the model had an immense knowledge difference compared to its predecessors, but also the semantic understanding of the natural language questions was raised significantly.

But GPT-4 was the one, in my opinion, that made all the difference. And the difference is in the decision-making process it can offer.

Now let me explain what I mean here. When I say decision-making in this context, I mean the actual decisions the model is capable of making to resolve your request in the best way possible (at least for the model).

In simple terms, what that means is the model can decide how to approach your question. A lot of questions are typically solved in a couple of steps. Now, GPT-4 can do it.

How GPT-4 decides to approach your question

Let’s say you’re analyzing the behavior of customers that recently signed up for your product. There're always going to be people that behave… very strange.

They can purchase a subscription in 1 second or they can be inactive for one month, then come back and get an annual membership.

The key here would be consistency. Is there a pattern in people doing that? If not - you shouldn’t factor them in your end analysis; otherwise, the end results will be skewed.

In statistical terms, these people are called outliers. And to determine who are outliers and who are not, you’d typically go through a couple of steps (I am using z-score method in this example):

  1. Calculating the historical mean

  2. Calculating the Standard Deviation

  3. Computing Z-score and pinpointing the customers beyond the threshold

This is a very common thing analysts would typically do while providing you with some report.

Now if you want to do this using Tableau, it would be really hard, since you will have to create a lot of custom properties.

And before GPT-4, any existing Text to SQL models, unless trained on this specific kind of question, would really struggle to generate it from scratch, since the query is insanely huge.

But GPT-4 completely changed the game in this regard. Now it can outline for you what should be done to spot the outliers and go do it step by step. As if it was a human analyst doing the work.

And since the queries, still being big, are logically not very complex - you will be able to get the answer in probably 2 minutes.

What is going to happen to the BI industry

Those of you working closely with BI tech might be wondering: “So you’re saying PowerBI, Tableau, and others will fade away while AI will conquer the world?”.

Practically, yes, you will not find these tools in the way they work right now in 5-7 years. Will they be replaced? Partially.

Mostly because these companies are not stupid and are already starting to adopt Text to SQL to slowly move away from the old user interface architecture.

Why did I say partially then? Well, that’s because since they have been optimized and developed with the old architecture in mind for the last 20 years, it would be really hard to adapt it to the whole new reality of words.

So most likely they’re just going to plug the LLM model on top of their stack and train it to operate with it. Which will probably work, but to the extent of how limited their existing stack is.

I doubt if the above example with outliers will ever be easily solved by any major BI tools without custom pre-configurations.

So in my opinion, there’s going to be a new wave (which is already happening) of Text-to-SQL analytics alternatives that will try to push out the big fish in the market. Whereas companies like Microsoft, Google will be trying to integrate AI in their analytics software which is going to take a long time.

The question is if any of them will be able to do it in the way that it does actually work?

So what can Text to SQL actually do right now?

Providing Answers Out of the Box

Right now, models can give you pretty good quality answers out of the box while feeding them the information about your company and your database correctly.

It does still require some minimal mapping, but it’s even hard to call it mapping. All you should do is to explain to AI the columns in your database that are unique to your business (like abbreviations) so then when you use these words, the model can “map” them to the actual fields.

This allows the models to provide answers out of the box without the need for extensive documentation or hard-coded definitions.

Multi-Steps Execution

As we talked about earlier, the advanced models can now generate a strategy for the upcoming steps they will need to undertake to get the data you’re asking for.

This also engages the user in the loop, so then it’s easy to follow and understand what AI is trying to do, adjusting it on the way.

Surprisingly, but as we discovered, the majority of mistakes AI do when trying to generate you a query is in misunderstanding your initial request. So if you can spot and correct it, the results will be really accurate.

Multi-steps execution of the requests facilitates exactly that.

Generating Charts

It goes without saying that if it can find data for you, it can present it in any way you can imagine. GPT-4 data analysis module is really good at plotting different kinds of charts, so you realistically will not need to do it in Excel anymore.

If you ask me how to deal with the charts that you want to keep on a dashboard and update every day, you can ask the model to generate you a chart based on the parameters provided and then simply save these parameters as well as a query for future users.

Semantic Layer

What if you say you already have so many queries created in different software that it would be nuts to redo it in a new tool even with AI helping you along the way.

That concern is tackled by the thing called a semantic layer, or what you can call a knowledge vault of your model.

Basically, you can import all of the queries you have in your SQL clients into it alongside with definitions of what each query means for the model to then use it in the future requests.

Let’s say you have a custom query and formula for lifetime value calculation of your customers. You can achieve the same by guiding the AI to it every time which becomes pretty annoying at some point.

Or you can define it in your semantic layer and then every time you’d ask “What’s the LTV of Q1 cohort?” kind of question the model will not try to generate anything from scratch but rather use your example to only adjust it based on the conditions in your request (e.g., timeframe or segmentation).

How Do You Know All That and Where Can We Try It?

We know that because we’ve been working on this the whole of 2024.

We tried so many things to see what works and what doesn’t to finally come up with a version that can actually do all the things described above. And this is probably the only product that can do it for now 🤞

A week ago, we released a free tool for everyone to use that can turn questions into queries only by knowing your schema.

We don’t charge for it, don’t even ask for an email. Only because we want to show you that it actually does work. You can try it here: datalynx.ai/text-to-sql.

If you want to try the product that can do everything mentioned above, you can sign up for the waitlist, which already has the trust of over 1000 people: Join the waitlist.

And if you have any suggestions or would like to help us move this project further, please do feel free to reach out at aleks@datalynx.ai.

I hope this article made you look at the problem of no-code analytics from a new perspective and raised a curious spirit to go give Text to SQL solutions a shot.

And even if it’s not our tool, I’m really pumped to see you joining this whole motion.

You rock! 😎

Here's some other great resources on this topic

  1. BirdSQL - Model to Human Level comparison benchmark

  2. SQLCoder-70b - New model that beats GPT4 based on CodeLlama-70B

  3. Spider - Another great benchmark list

What's Text to SQL? Is it still a thing? No, you can't replace nobody, the tech is not there!

All that and many more we've been told for the last year. Yet our opinions remain very polarized.

In this article, I'll show you why we believe that you can and probably should prioritize adopting AI in your organization instead of hiring more analysts.

We spent 12 months building a text to sql tool that can do 70% of the work most analysts do nowadays. And not only us, hundreds of others work 100-hour weeks to push this bar further and further.

This is happening, whether you like it or not.

Today you'll learn:

  • What Text to SQL is all about

  • How LLMs completely flipped the text-to-SQL movement

  • What is going to happen to the BI industry

  • What current solutions can exactly do

And so much more! Let's dive in… but first

TLDR - You Can’t Replace Them In One Night

I know you might be here thinking, "yeah, obviously you can't," but just hang on for a second. The idea here (which took us a year to understand) is that you CAN replace them, it's just not instantaneous.

But we tried. I mean, we really did think that when ChatGPT came out, it was the day. Not even close.

If you simply take an OpenAI API key, as many might think, and connect it to the front-end chatbot and try asking it some analytical questions (also providing your database information), in 50% of the cases, it will not work.

Why? Because:

  1. Your database is too big and can't fit in the prompt.

  2. The user’s query is too complex, and AI can't fit it all in one query.

  3. You're using GPT-3.5, which simply doesn't work.

  4. Your database structure is so messed up that AI has no idea how to translate it into relevant code.

  5. You don't understand how AI works, so the way you phrase your questions will make them never be answered.

  6. Hope you get the point.

Some of the fixes to the above require just extra coding. Some require extra training. But some require extra something that we had no idea about.

But in the end, if you connect it all together... it might work. And we see it's working pretty darn well!

What Is Text to SQL And How Does It Work?

Text to SQL is a natural language processing technology that converts plain text queries into structured SQL queries. It works by analyzing the text, identifying key elements such as entities and relationships, and transforming them into SQL.

The idea behind it is bread and butter - how to make people who do not know how to write code achieve things people who typically know how to write code do achieve.

And no, it's not new by any means. The BI (Business Intelligence) trend started in the 1950s with exactly the same idea.

The issue is it hasn't moved anywhere from there for the last 70 years. Obviously, there were some improvements and breakthroughs, but the core concept remained the same - translating business requests into SQL code using a user interface.

The whole BI trend was progressing, not because it was really innovating, but because the tech behind it was becoming better and better. Computers became faster, processing time shorter, interface cleaner.

But in a nutshell, it remained the same interface translation into generated sql queries. And nobody could do anything about it as it was a reliable and honestly single reliable and scalable way to receive insights from the dataset in a low latency way.

And as you know, different strokes for different folks, so it didn’t take too long for the alternatives to come up.

Text to SQL handles that by using LLMs to translate user requests into SQL code, so you don’t have to spend hours configuring the dashboard in Tableau.

It’s as simple as asking a question in english and getting an answer… or not?

How LLMs completely flipped the text-to-sql movement

When OpenAI released GPT-3.5, it changed the whole scene for everyone working on the text-to-SQL technology.

Not only did the overall accuracy of queries become better as the model had an immense knowledge difference compared to its predecessors, but also the semantic understanding of the natural language questions was raised significantly.

But GPT-4 was the one, in my opinion, that made all the difference. And the difference is in the decision-making process it can offer.

Now let me explain what I mean here. When I say decision-making in this context, I mean the actual decisions the model is capable of making to resolve your request in the best way possible (at least for the model).

In simple terms, what that means is the model can decide how to approach your question. A lot of questions are typically solved in a couple of steps. Now, GPT-4 can do it.

How GPT-4 decides to approach your question

Let’s say you’re analyzing the behavior of customers that recently signed up for your product. There're always going to be people that behave… very strange.

They can purchase a subscription in 1 second or they can be inactive for one month, then come back and get an annual membership.

The key here would be consistency. Is there a pattern in people doing that? If not - you shouldn’t factor them in your end analysis; otherwise, the end results will be skewed.

In statistical terms, these people are called outliers. And to determine who are outliers and who are not, you’d typically go through a couple of steps (I am using z-score method in this example):

  1. Calculating the historical mean

  2. Calculating the Standard Deviation

  3. Computing Z-score and pinpointing the customers beyond the threshold

This is a very common thing analysts would typically do while providing you with some report.

Now if you want to do this using Tableau, it would be really hard, since you will have to create a lot of custom properties.

And before GPT-4, any existing Text to SQL models, unless trained on this specific kind of question, would really struggle to generate it from scratch, since the query is insanely huge.

But GPT-4 completely changed the game in this regard. Now it can outline for you what should be done to spot the outliers and go do it step by step. As if it was a human analyst doing the work.

And since the queries, still being big, are logically not very complex - you will be able to get the answer in probably 2 minutes.

What is going to happen to the BI industry

Those of you working closely with BI tech might be wondering: “So you’re saying PowerBI, Tableau, and others will fade away while AI will conquer the world?”.

Practically, yes, you will not find these tools in the way they work right now in 5-7 years. Will they be replaced? Partially.

Mostly because these companies are not stupid and are already starting to adopt Text to SQL to slowly move away from the old user interface architecture.

Why did I say partially then? Well, that’s because since they have been optimized and developed with the old architecture in mind for the last 20 years, it would be really hard to adapt it to the whole new reality of words.

So most likely they’re just going to plug the LLM model on top of their stack and train it to operate with it. Which will probably work, but to the extent of how limited their existing stack is.

I doubt if the above example with outliers will ever be easily solved by any major BI tools without custom pre-configurations.

So in my opinion, there’s going to be a new wave (which is already happening) of Text-to-SQL analytics alternatives that will try to push out the big fish in the market. Whereas companies like Microsoft, Google will be trying to integrate AI in their analytics software which is going to take a long time.

The question is if any of them will be able to do it in the way that it does actually work?

So what can Text to SQL actually do right now?

Providing Answers Out of the Box

Right now, models can give you pretty good quality answers out of the box while feeding them the information about your company and your database correctly.

It does still require some minimal mapping, but it’s even hard to call it mapping. All you should do is to explain to AI the columns in your database that are unique to your business (like abbreviations) so then when you use these words, the model can “map” them to the actual fields.

This allows the models to provide answers out of the box without the need for extensive documentation or hard-coded definitions.

Multi-Steps Execution

As we talked about earlier, the advanced models can now generate a strategy for the upcoming steps they will need to undertake to get the data you’re asking for.

This also engages the user in the loop, so then it’s easy to follow and understand what AI is trying to do, adjusting it on the way.

Surprisingly, but as we discovered, the majority of mistakes AI do when trying to generate you a query is in misunderstanding your initial request. So if you can spot and correct it, the results will be really accurate.

Multi-steps execution of the requests facilitates exactly that.

Generating Charts

It goes without saying that if it can find data for you, it can present it in any way you can imagine. GPT-4 data analysis module is really good at plotting different kinds of charts, so you realistically will not need to do it in Excel anymore.

If you ask me how to deal with the charts that you want to keep on a dashboard and update every day, you can ask the model to generate you a chart based on the parameters provided and then simply save these parameters as well as a query for future users.

Semantic Layer

What if you say you already have so many queries created in different software that it would be nuts to redo it in a new tool even with AI helping you along the way.

That concern is tackled by the thing called a semantic layer, or what you can call a knowledge vault of your model.

Basically, you can import all of the queries you have in your SQL clients into it alongside with definitions of what each query means for the model to then use it in the future requests.

Let’s say you have a custom query and formula for lifetime value calculation of your customers. You can achieve the same by guiding the AI to it every time which becomes pretty annoying at some point.

Or you can define it in your semantic layer and then every time you’d ask “What’s the LTV of Q1 cohort?” kind of question the model will not try to generate anything from scratch but rather use your example to only adjust it based on the conditions in your request (e.g., timeframe or segmentation).

How Do You Know All That and Where Can We Try It?

We know that because we’ve been working on this the whole of 2024.

We tried so many things to see what works and what doesn’t to finally come up with a version that can actually do all the things described above. And this is probably the only product that can do it for now 🤞

A week ago, we released a free tool for everyone to use that can turn questions into queries only by knowing your schema.

We don’t charge for it, don’t even ask for an email. Only because we want to show you that it actually does work. You can try it here: datalynx.ai/text-to-sql.

If you want to try the product that can do everything mentioned above, you can sign up for the waitlist, which already has the trust of over 1000 people: Join the waitlist.

And if you have any suggestions or would like to help us move this project further, please do feel free to reach out at aleks@datalynx.ai.

I hope this article made you look at the problem of no-code analytics from a new perspective and raised a curious spirit to go give Text to SQL solutions a shot.

And even if it’s not our tool, I’m really pumped to see you joining this whole motion.

You rock! 😎

Here's some other great resources on this topic

  1. BirdSQL - Model to Human Level comparison benchmark

  2. SQLCoder-70b - New model that beats GPT4 based on CodeLlama-70B

  3. Spider - Another great benchmark list

Stop using chatGPT for SQL today

Think about the last time you had a business question. How long did it take to answer it?

Stop using chatGPT for SQL today

Think about the last time you had a business question. How long did it take to answer it?

Stop using chatGPT for SQL today

Think about the last time you had a business question. How long did it take to answer it?

Copyright © 2024 Docugenie, Inc.

Copyright © 2024 Docugenie, Inc.

Copyright © 2024 Docugenie, Inc.