Why we built OpenDash with unlimited context

Most data questions don't fit in a chat window — they span schemas, definitions, and weeks of tribal knowledge. Here's how we're trying to fix that.

If you've ever asked a chatbot a question about a customer cohort, watched it produce a confidently wrong answer, and then spent three hours tracking down which of your five definitions of "active user" it picked — you've hit the context problem.

A typical analytical question touches dozens of files, hundreds of table columns, and conversations with people who've been there for years. None of that fits in a chat window. Not a large one, not even a 200k one.

The thing most tools get wrong

We tried it the normal way first. We built a data assistant with a fixed context window, watched it forget things, and shipped it anyway because that's what everyone was doing. The result was the same as everyone else: it gave answers that looked confident and quietly disagreed with the source-of-truth dashboard two floors up.

So we pulled it apart.

How we think about it

Our approach has three ideas behind it, none of them original but all of them mostly missing from current tools:

1. Pull the source of truth, not the LLM's guess. Tables, schemas, code, definitions — load them from where they actually live, not from a snapshot the model was trained on.
2. Show every step. If OpenDash used a CTE that joined three tables and applied a 14-day filter, that's all visible, clickable, and editable.
3. Question everything that disagrees. If two figures for the same thing don't match, we surface that instead of picking one quietly.

What changed for our users

In the past three months:

• Time to answer for a "what was Q3 revenue by region?" question fell from ~22 minutes to ~3.
• Adoption moved from a few analytics engineers to a few hundred analysts, PMs, and operators.
• The most common thank-you message: "It actually agreed with our dashboard."

What's next

We're working on a few things on top of this foundation:

• Memory. Per-team memory that learns from past analyses, definitions, and corrections — without leaking between teams.
• Sub-agents that prove their work. Today you can launch ten parallel investigations. Soon they'll show their reasoning and trade-offs side-by-side.
• Sharing that doesn't lose provenance. Share a result with a Slack message and the chart arrives with the query and grain attached, so nobody quotes a number out of context.

If you've ever been burned by an AI tool confidently producing the wrong number, we'd love for you to try OpenDash. OpenDash is open source and free to use — we'd rather earn your trust than your data.