“Our accuracy drifted from 95% at launch to 65% over a month.”
Keeping your data agents accurate is a never-ending job.
-
You can’t anticipate every question.
“Do trial accounts count as active?”
-
A change can break a good answer.
“Are we using the new refunds table?”
-
The next team brings more context to maintain.
“What’s our customer retention rate?”
Cassis is the context engine behind your data agents.
Keep context current at scale.
Questions and data changes keep coming. Cassis groups recurring gaps and prepares fixes. Your team reviews one change for all the questions behind it.
Example: questions from Sales, Finance, Growth and Exec about active customers become one fix, "Active customers exclude trial accounts", assigned to Léa M. A dbt table dropped and a new one created become "refunds renamed to refunds_v2", assigned to Tom R. Two churn questions become "Churn counts accounts, not seats", assigned to Inès K. Each fix passes tests and evals and is ready to publish.
2,418 questions and 37 data changes this week
Reliable answers everywhere.
Active customer
Excludes trial accounts.
- Tables
customersorders- Metrics
- Active customer count
Active customer growth
+8.4% this quarter.
Build from what you already have.
- Schemas
- dbt models
- Dashboards
- Docs
- Semantic layer
- Catalog
Context review
87% auto-documented
- 10 business domains
- 155 metrics
- 532 tables
13% pending review
Built for scale.
Improves with use.
Capture knowledge when people have a reason to share it.
Cassis proactively asks users for missing context, feeding their answers into the feedback loop.
Keep large context usable.
Cassis gives your agent only the domains and definitions a question needs, so answers stay sharp as context grows.
Focus on decisions.
Cassis manages evals, CI, reviewed publication, permissions, and everything that breaks as you scale.
Augment your data stack.
Keep your agents and warehouse. Put your existing definitions to work.
Your agents
Claude
ChatGPT
Copilot
- Custom agent
Cassis
Maintained context
+ answering
Your data stack
Your semantic layer
Your warehouse
Snowflake
Databricks
BigQuery
PostgreSQL
Your Git repository
GitHub
GitLab
Manage your context as code.
-
Edit your repo
› Apply Cassis's fix active_customers.yml
Claude Code, Cursor or Codex writes the file.
-
Test the fix
$ cassis context test 11,260, trials excluded
Check it answers the questions that raised the issue.
-
Open a pull request
context check eval suite 12/12
CI runs the evals so nothing else regresses. Your team reviews.
-
Merge to publish
#23 merged v42 published
The next answer uses it.
Frequently asked questions
What data does Cassis access?
You choose between two modes.
- Metadata only, no warehouse connection: Cassis uses your schema, dbt project, documentation and questions to generate SQL. Your existing tools run it in your warehouse. Row data stays in your environment.
- Read-only warehouse connection: Cassis executes the SQL through a dedicated read-only role, limited to the datasets you authorize. No warehouse writes.
In both modes, Cassis sees the questions your assistant sends it, not the rest of the conversation.
What if we already have a semantic layer?
Keep it. Cassis reads its definitions, Snowflake semantic views included, and runs queries through it when connected to your warehouse. Cassis adds two things:
- More context: business concepts, rules and exceptions by domain, which table answers which question, and known data limitations.
- The engine around it: one MCP server for all your agents, and maintenance. Questions surface what’s missing, Cassis proposes the fix, evals test it, and your team reviews it as a pull request in your own Git repository.
Can you build our first context for us?
Yes. We build the context for one business domain from what you already have: schema, dbt models, dashboards and docs.
Your team settles the definitions we can’t infer, and we test the context against real questions from that domain. Then your team maintains it with Cassis, reviewing each fix before it’s published.
What does our team need to review?
The fixes Cassis proposes and the business rules it can’t settle alone. Your team can edit or decline any fix. Cassis groups the questions behind each gap, so one review covers all of them.
Nothing is published without your data team’s approval. Each change is versioned, tested with evals and reviewed in Cassis or as a Git pull request. Edit rights can be scoped per domain.
Our BI tool already connects to agents. Why Cassis?
Your BI tool’s agent only knows your BI tool’s model, and your people already ask Claude, ChatGPT or Copilot too. Cassis gives every agent the same approved context, including the rules your BI model doesn’t hold.
Your BI tool stays. An agent can still build the chart or dashboard there, with Cassis supplying the definitions behind it.
Do we still need Cassis if we have a data catalog?
A catalog records what exists, and it ages fast. Cassis learns from your people’s questions what’s missing and keeps the context current.
It starts from your catalog: Cassis imports its descriptions and glossary, and can stay in sync if you need it. Questions and corrections then reveal gaps and conflicts, and Cassis prepares updates for your team to review.
Why not build this ourselves?
It’s the best way to start, and you can go all the way if you have the resources. What we see is that it gets hard at scale:
- Analysis: reading thousands of questions consistently to see what fails, then grouping failures by cause and ranking them so your data team isn’t buried in tickets.
- Learning curve: we see more context gaps, and more ways to fix them, across deployments. Every customer benefits.
- Plumbing: evals, CI, permissions and governance, all maintained as you grow.
Get your data ready for analytics agents, and keep it that way.
Bring one domain and we'll go through it with you: what your agent would get wrong today, which definitions your team has never settled, and what keeping them current takes.
Notes on data, meaning, and context
LLM analytics benchmark: the best model for your analytics agent
We compare eight LLMs on analytics questions to choose our agent’s model, measuring answer quality, reasoning effort, cost and the effect of our harness.
Your data context should be debuggable
How a tree of business domains makes analytics context easier to debug, maintain, and retrieve without overloading the agent.
How to assemble context for analytics agents from your existing assets
Your schema, dbt, query logs and docs already hold much of the context an analytics agent needs. Here's how to recover it without inventing what isn't there.