ClaireGz

Yes exactly how I felt - I wanted to keep control on the tooling + on what's inside the context so Git felt like the most open tool to do this

OpenAI recently shared how their internal data agent works: [https://openai.com/index/inside-our-in-house-data-agent/](https://openai.com/index/inside-our-in-house-data-agent/)

What interested me most wasn’t the model itself, but the surrounding system design: automated context retrieval, evaluation loops, and memory that improves the agent over time.

I’ve been experimenting with recreating a similar setup, but with a different goal: making the configuration more accessible for any company.

The result is an open-source YAML + Markdown framework where you define context sources, tools, and behavior explicitly instead of writing Python code. The idea is to make agent context easier to reason about, version, and iterate on, especially for data teams.

Repo: [https://github.com/getnao/nao](https://github.com/getnao/nao)

Would love feedback from people who have tried deploying analytics or data agents before.

A lot of agent failures I’ve seen come from small details getting dropped when context is summarized or compressed.

How do you make sure the densification step doesn’t remove something that ends up being important for the task?

Interesting direction.

One thing I keep seeing in practice is that “memory” problems are often less about storage and more about structure + retrieval strategy.

Vector search helps sometimes, but for a lot of agent workflows we’ve had better results with explicit context organization (files, metadata, rules) rather than semantic similarity alone.

Curious how you’re thinking about memory updates over time — append-only vs rewriting summaries?

This is super helpful — most writeups skip over the actual communication steps, so seeing the All-to-All flow laid out makes it much clearer.

Curious from your experiments: at 1M+ context, does communication start dominating vs compute?

I keep seeing cases where bigger context windows are technically possible but don’t translate into better results unless the context is very structured, so I wonder where the real scaling limit ends up being in practice.

As an update, we now have SSL connections available

As an update, we now have SSL connections available

Here is our contact page, feel free to contact us and we'll hep you setup: https://docs.getnao.io/docs/support/support Also, we'll be adding a new dbt onboarding flow tomorrow!

It's Databricks, Iceberg and Redshift, which on the first survey that we did were the most asked. But as per this post and a broader audience it appears SQLite at least win! We'll add also SQLServer in the list

Thanks for your feedback, we'll add SSL in the connection soon

Probably in a few months. For now we're focusing to make the experience great for a restricted number of warehouses. But you can reach out by email and we'll keep you updated

Thanks for sharing. I like the view you built to visualize the profiling of your data, I think that's indeed key to understand your data.

The content of the data is never going to the models, unless you specifically grant access. Our server only store embeddings of your codebase and data schema. The content of the data is only accessed locally by your computer. When you ask our agent to run a query for you, it will execute it on your warehouse, then ask you for permission to read the result. If you don't allow it, you'll be able to preview it locally without sending it to the LLM. The enterprise version is for your if you want to be sure the prompts and context you send to nao don't go through a public LLM endpoint and get trained on.

Thanks for suggesting, we should set this up!

Thanks for your kind message — and for helping us make nao better!

Glad you liked it!

Yeah at the moment the Tab is made to work the best with raw SQL (either pure SQL files or in a string).

But, if you use the chat/agent you can explain that you're using Kysely and give the warehouse context he will probably handle this.

I did not know Kysely, but from the gif on the project landing it looks like the autocomplete is great? It's different than a tab I agree tho.

Not yet, but we are willing to develop these specific connectors on ask. You can reach out!

Just one question what makes you pick Exasol rather than going with an open-source warehouse tech (e.g. Clickhouse or a lake with Trino)?

We heard of Julius a lot, but did not know about Cipher42, there are a few others folks around. We feel there is a pain and also data teams are a bit abandoned at the moment when it comes to work using AI so makes sense. Curious to get a feedback about your journey building cipher, did you stop working on it?

Oh yes! Should be done fairly easily, we have DuckDB coming in the next release, we can also add SQLite. You use SQLite to develop locally I guess?

we support Postgres (and DuckDB is coming very soon) so yes probably as Hydra is a mix of both, but I have to try it

Usually what we've seen is data people having notebooks/worksheets on the side with a bunch of manual SQL queries that they run to validate the data consistency. The process is highly manual, time consuming. Most of the time teams knows what kind of checks they want to run on the data to validate it, our goal here is to provide them the best toolbox to do it, in the IDE.

Tho, i'd say this is like when writing tests in software, you can't catch everything the first time (even when going 100% code coverage), especially in data when most of the time it breaks because of upstream producers.

It will still require live observability tools monitoring data live in the near future.

Thank you!

We say "data vibing" so it feels unique to the data community! But in all seriousness, this is already an issue, people are already asking ChatGPT (or Cursor/whatever else) to generate SQL for them, but the next steps do not exist, if you "vibe code" for data you want to have the easiest feedback loop you can get to check if the output is good, and that's what we are working on: identifying the downstream impacts in the IDE and proposing fixes, a table diff view, new UI/UX to test your outputs.

The goal for us is to be the best way to do data with AI.

Thank you!