Dex Semantics - Let your agent build analytics with 25% fewer errors
by•
Dex Semantics adds declarations to your analytical models, which capture row meaning, join nature, how numbers add up, and decisional assumptions.
Thanks to this, Dex scores 80% on ADE-bench with Claude Sonnet 5.5, compared to 73.3% without Dex (25% fewer tasks failed).
Declarations live in your models as code, not a third-party tool. Dex then helps your agent convert declarations into your Semantic Layer of choice.
Replies
Hey everyone, I think this quote neatly describes the problem we're solving.
It's acutely painful in the context of analytics, where one wrong (but plausible) number can steer a giant enterprise or a fast-growing startup in the wrong (and disastrous) direction.
And, of course, the problem's magnitude is amplified by the speed of development brought about by agents.
Thank you @fmerian and @gtmnerd!
@gtmnerd @marco_ciavarella What's the best way to structure semantic declarations for older datasets that lack proper foreign key constraints?
@deepaksingh09 great great question.
I'd start with this command, which lets you infer and measure joins:
This is useful for brownfield warehouses/databases where foreign key constraints may not be explicit, but can be inferred (and then made explicit after some cleaning with `dex transform`).
Then add declarations and run:
Once initial set up is done, these commands on a schedule help you evolve declarations as the data changes:
This entire workflow is very agent-native, lots of details to check, so let your coding agent do this work for you!
More info here: docs.exmergo.com/dex
Specific declarations docs, for you and your agents, is here on GitHub: https://github.com/exmergo/dex/blob/main/skills/transform/references/declarations.md
@marco_ciavarella Fully agree. A plausible but wrong number is the hardest bug to spot, and that’s the problem we care about most.
I like that you're keeping semantic definitions close to the actual code, i think it help optimize SQL queries as well.
@priya_kushwaha1 Definitely!
@priya_kushwaha1 thank you! keeping it next to the code was a big decision for us, glad it resonates
@emna_rahmouni best of luck.....
Hi Marco, congrats on the launch!
I'm a terminal nerd, and Dex sounds like it's really cool. I'm going to check it out.
@nathan_tarbert Awesome! We specifically built Dex as a CLI tool to have both people and agents be able to work with it. No MCP or login needed, just install the CLI (and skills, if you want to give your agent the guidelines on how to use Dex) and let it rip!
@nathan_tarbert a fellow terminal nerd! let us know how it goes 🙌
@emna_rahmouni, terminal nerds need to stick together :)
Let's goooooo!
@gtmnerd I ♥️ GTM
Really interesting to see a toolkit addressing both the intelligence and reliability sides of AI-driven analytics. Those two things need to go together.
@tehreem_fatima5 Very well put
@tehreem_fatima5 thank you, glad that comes through!
Is this tool for building knowledge graph to set up the context layer for AI agents in enterprise?
@vijayanandkailash 100%, Dex helps agents author and connect to 5 of the leading Semantic Layers in the space:
dbt Semantic Layer, Motley's SLayer, Cube, Snowflake Semantic Views, Databricks Metric Views.
Dex also authors Apache Ossie, the emerging standard for the Semantic Layer. We're part of their ecosystem! https://ossie.apache.org/ecosystem/
More semantic layers coming soon!
@marco_ciavarella Congrats for launch Team does it automatically suggest updates to the declarations when a database migration runs in CI?
@rohit62661 Thank you!
You can configure agentic workflows in your CI that leverage the `dex maintain` commands for this!
We're actually working on integrating this specific feature you mentioned in our own managed workflows: https://trydex.ai/enterprise
The YAML semantic layer is the part that caught my eye — most agent-built dashboards I've seen fail on inconsistent metric definitions, not on SQL. Curious how you handle it when two teams define the same metric differently, or when a definition drifts from what's actually in the warehouse?
@jcbuilds
`dex maintain` is all about handling this problem! https://github.com/exmergo/dex/blob/main/skills/maintain/SKILL.md
Interesting point. We recommend, at an organizational level, to define one semantic layer per vertical (finance, ops, internal metrics etc.) so that this doesn't happen.
Our enterprise offering is all about handling these types of problems at scale, with our agentic workflows: https://trydex.ai/enterprise
Really cool Dune inspiration demo :)
What does it exactly mean by living on my models as code? Is it a harness?
Congrats #makers!
How do you keep semantic declarations accurate when the underlying data model changes over time?