Show HN: Sourcetable – AI Spreadsheet and Data Platform

> For my particular use case, the charting aspect left a few things to be desired - once I grouped campaign donations by contributor, I could only see the first 10 rows in the AI response, with no option to expand the output.

Insert it as a table on the page (you should see a button), it will then print the whole table result from that query into the spreadsheet. Also, you can check the SQL first and validate it, then print to table after that.

Try a few million rows and see what happens!

Can Office 365 Copilot do this too?

Minus the AI part tools like this have existed for decades.

And companies are not dumping their SaaS tools and switching to them en masse.

Because (a) data silos have dramatically increased pushing dreams of a unified data schema out of reach, (b) technology stacks have become far more complex necessitating tools like Datadog and (c) competition is stronger than ever meaning that skimping on paying for tools like MixPanel is often short sighted and counter productive.

Companies like this will do fine and there will be always be a demand for them especially in the SMB space. But there simply isn't the business value in bringing a lot of analytics and observability in-house in almost all cases.

Agree. A general thesis I have is that the API-ification of the web fragmented business information, and with every new SaaS tool we fragment our company's data further. The trend at all company sizes is to be increasingly analytical, but for SMBs it's too hard to get access to your data (mainly due to technical limitations). So it makes sense to centralize data somewhere, and we think that somewhere is inside the data tool that everyone actually uses: the spreadsheet.

Many other advantages of this data centralization too. Data + spreadsheets + compute is a nice application base for agents.

At risk of poking the bear, they should have done this decades ago. Except for LLMs they have had everything they needed to bundle this stack into a single product solution; this would be much better for users.

And yes! We're definitely of the opinion that as a startup we can outcompete the two trillion-dollar death stars when it comes to product experience. AI is a platform shift!

When I was in there in Microsoft Research, our team was working on related efforts. But, yes, while pieces have shipped into products, MS never released a complete solution at the time.

Some links that might be of interest:

- Table semantics: https://www.microsoft.com/en-us/research/project/table-inter...

- Entity semantics (video): https://onedrive.live.com/?authkey=%21AMIdbT4yVFaw2Kk&cid=A6...

- Natural Language in Spreadsheets: https://www.microsoft.com/en-us/research/project/gridbook/

Actually Microsoft do now have Copilot and Python in Excel recently released last week. Maybe a bit slow.

Category Comparison (table-based solutions): "How are you different than Retool/Airtable/Coda/Notion/Zapier Tables, etc."

The primary difference vs table-based solutions is that Sourcetable is a spreadsheet in the common sense of the word, similar to Excel and Sheets. We have A1 notation and cell-based referencing. This is what most users expect, and this flexibility/familiarity has a big impact on the breadth of users and use cases within a team.

The formula referencing system of these table-based solutions is usually very limited both to columns/rows (not cells), and is a set of SQL-based queries which are much more limited than that 500+ formulas and functions spreadsheet users commonly expect.

Retool specifically: I tend to think of Retool as a lightweight custom-ERP software system, whereas Sourcetable more like Excel + PowerBI + Data Warehouse, so we will generally be much stronger for reporting and analysis. We definitely have some overlap in potential users since technical operators should like us both. FWIW - Retool is an excellent product.

Hi I'm Andy, Cofounder & CTO @ Sourcetable.

We use a heavily modified licensed engine that prevents us from open sourcing everything (for now). We have plans to open source our agentic/plugin framework, and other parts of the system. We also have a strong ethos of contributing back to open source where we can (contributed back to Arrow, DuckDB etc.).

I'd also add that while everyone knows how to use and work with spreadsheets, we also provide a SQL layer on top that you can use to query data sources as an advanced user (we developed a nomenclature to work within sheets/across sheets/files/our data-warehouse). This allows more technical users to work side-by-side in the same environment as non-technical users without crossing pythonic or reporting boundaries.

On top of this, the AI assistant can answer most of the questions you might have of all this data.

I think as ML gets more sophisticated, we will in general need to be less technical. The "tooling" might even disappear, but we will still need something to communicate important data centric decisions. Whether you like it or not spreadsheets are the foundation of human research and operations and have been for thousands of years, and I feel humanity will need less complicated "tools" and we will keep to our roots.

Agree with you, and we're definitely trying to thread the needle!

We're generating the SQL to answer natural language questions, so folks can just get answers and results tables if that's all they need, with the option for power users to fiddle with the SQL either directly or via a query editor GUI.

There's a ton of use cases for working with unstructured and semi-structured data and that's coming down the pipe!

This is 100% the correct insight in my experience.

TL;DR, most technical people massively overestimate the technical / data abilities of regular spreadsheet users. We find simple use cases are best, and with each new LLM release the UX around more complex data improves significantly.

The reason we chose to build as a full-blown spreadsheet instead of just a table-based solution was that we saw that most people want the flexibility of a regular spreadsheet, but access to their (structured) business data. Table-based solutions wedge you into AI and you can never get out of that.

Thanks for reporting! Taking a look now.

A very common use case we see is SMBs having outgrown their spreadsheet but not wanting to move to a full-blown BI tool. They want the power, but not the change in interface/medium.

I didn't go into details above but a nice thing is that we leverage cloud compute and storage, so you can query billion-row data in sub-second time. (Courtesy of Duck!)

Hi - we encrypt everything at rest and use metadata to improve LLM performance.

We don't yet have enterprise-grade data permissioning or compliance certificates like Soc2. Those will come in time.

The underlying datatypes we support in our data-warehouse support 3d and 4d data. So we can do vector queries on these and do transformations over different spaces. I think given what you need we can put your data in our data-warehouse, and then present it to the older engineers in an excel format with 3d plotting. We might want to chat about the details though, give me a holler at andrew@sourcetable.com

Yes actually! My cousin is a mining engineer so I spent a bunch of time playing around with mining data during testing. Turns out all New South Wales government data is public. Right now you can talk to any CSV or database using LLMs. I've also played around with a bunch of marine biology datasets too!

(p.s. I think Andrew, CTO, is going to jump in here as he has more experience in this space.)

Thank you!

You may see that we try to suggest follow-up questions or question improvements where we think better context-in will result in a better result-out.

Curious what will happen if you modify the question to be more explicit?

I have seen that PMs and data-trained folk tend to be very articulate in asking for exactly what they want and that tends to lead to significantly better LLM responses.

Thanks mate!

Fivetran will build a Streak integration if you bring a customer who will use it (sometimes as little as one): https://fivetran.com/docs/by-request-program

Mostly Fivetran, a little Airbyte, and a few custom integrations. Would love to add Streak (can you get it into Fivetran? We can usually crank those integrations out within an hour.)

If you found any bugs please let me know! andrew@sourcetable.com

Yes we don't yet have the full auto-suggest magic that Sheets offers, but you can click-drag for auto-complete the same way Excel offers.

We released Sourcetable today with the AI chatbot & AI data analysis features, but a very limited cell-based AI (only "summarize" and "fix formula"). We'll be releasing a big AI-based magic-autofill solution in the coming weeks.

Wherever possible, the chatbot output is deterministic, in that to answer a query, we're realtime generating and running code or SQL against your data. Our LLM orchestrates that, and finally evaluates whether the output correctly and adequately answers the question.

We also extensively use synthetic data and examples to guide and constrain our models.

Another way we're ensuring good-quality output is to ensure good-quality _input_ -- by enriching the detail and specificity of the user's question, and asking the user to disambiguate when we determine the question is too broad.

Would love to but unfortunately there are pieces we can't open-source for various reasons. We'll open source bits and pieces over time, and generally are excited to start blogging about AI & technical learnings now that the product is out of stealth mode.

Small plug for the analytics tracker we are using which Andrew (CTO) built and is open source: https://github.com/sfproductlabs/tracker

We use our own implementation of function calling orchestrated by chain-of-thought. The CoT allows us more granular control over the function calls, rather than zero-shotting and hoping the LLM selects the right functions.

Thankyou!!!

Yes definitely. We wrote a multi-step reasoning engine similar to o1. Depending on the route and depth in the chain of response, prompting techniques vary. The LLM router was pretty important to get right. On input: synthetic question enhancement has yielded positive results. Asking the right question is subtly important, and we find it is sometimes better to just slow a user down and help them improve their question instead of relying on multi-step reasoning from the get go. On output: checking answers via mixture of experts and other techniques is often necessary -- "truthful but wrong" answers are where some tricky gotcha's lie. Eval is hard.

One under-discussed feature of fast databases is that they are especially necessary in an agent-centric world. E.g. if you're running recursive SQL to pathfind towards an answer, it has to happen with minimal latency otherwise you break the user experience. Our interface looks like a spreadsheet and users mentally benchmark against spreadsheet-latency speed. They won't accept 20 second query response times they might get from their data warehouse.

Thank you!

You can't post personal attacks to HN. We ban accounts that do, so please don't do it again.

We detached this subthread from https://news.ycombinator.com/item?id=41597837.

I'll just speak up so this comment isn't left unchallenged. I'm friends with the folks on the Sourcetable team and I think they're good people, including the CTO.

[flagged]

Thanks _hfqa! We think there's massive potential here. It's a big platform shift, and spreadsheets weren't really impacted by the mobile or cloud compute waves, so it's a space long-overdue for disruption. (The last shift was back when Google Sheets took spreadsheets to the browser 17 years ago!!)