Conversation UI on Data
A specification for chat-with-your-data where the model picks certified metrics and never writes SQL.
Active · join meData Engineering
1 · Purpose
This is a specification, not an implementation: the right architecture for asking questions of data in plain English. It covers two settings, data governed in Unity Catalog on Databricks, and data in PostgreSQL or ClickHouse on a virtual machine. Its core rule is that the model selects a certified metric and dimensions, and a deterministic engine writes the query. It is for teams planning a conversational analytics product who want answers that match their dashboards.
2 · Business benefits
Trustworthy numbersChat and dashboards read the same semantic layer, so they always give the same answer.
Less reworkWriting the metric contract first stops gold tables being built in the wrong shape.
Clear tool choiceThe spec explains which semantic layer fits each setting and why.
Lower risk at launchShipping dashboards before chat exposes definition bugs to a forgiving audience first.
3 · Status and how to join
Active · join me. Under way: pick a task below, or tell me what you'd like to learn and I'll suggest one.
Good first tasks this week:
- Write twenty real business questions as metric, dimensions and grain.
- Review one track document and list any claim without a source.
4 · Requirements
What it must do
- Define what a right answer means in testable terms.
- Give a layered reference architecture from raw data to chat.
- Correct four common medallion assumptions, such as keeping gold atomic rather than aggregated.
- List what a semantic model must declare: grain, measures, dimensions, join paths, synonyms and owners.
- Map the architecture to each track's technology and compare semantic layer options.
- Set the order of work, from metric contract to golden questions to dashboards to chat.
How well it must do it
- No code is generated in this repository by design.
- Every claim about accuracy cites a public benchmark or source.
- Answers must be fast and cheap enough to ask a hundred times a day.
- Governance: access rules apply the same way in chat and dashboards.
5 · Architecture, very high level
6 · Tech stack, per platform
| Component | Databricks | Open stack | Azure | AWS |
|---|---|---|---|---|
| Clean data layer | Delta tables in Unity Catalog | dbt Core on PostgreSQL or ClickHouse | — | — |
| Atomic gold model | Delta star schema in Unity Catalog | dbt Core star schema | Azure Databricks or Microsoft Fabric | Amazon Redshift |
| Semantic layer | Unity Catalog Metric Views | Cube Core | Power BI semantic model | — |
| Pre-aggregation cache | — | Cube pre-aggregations | — | — |
| Dashboards | AI/BI dashboards | Apache Superset | — | Amazon QuickSight |
| Conversation UI | AI/BI Genie | Open-weight model calling Cube over MCP | Fabric data agent | Amazon Q in QuickSight |
7 · What needs to be done
Specification 5 of 5 done
| Task | Status | Size | Skill |
|---|---|---|---|
| Requirements in testable terms | done | M | requirements |
| Reference architecture and four corrections | done | M | architecture |
| Databricks track and self-hosted track | done | M | data platforms |
| Semantic layer options compared | done | M | analytics engineering |
| Decisions, open questions and risks | done | S | architecture |
Prove it on real data 0 of 3 done
| Task | Status | Size | Skill |
|---|---|---|---|
| Twenty real questions as metric, dimensions and grain | todo | M | analytics |
| Golden question set with human-checked answers | todo | M | evaluation |
| Apply the spec in a build (see AskSherlock) | doing | L | data engineering |
8 · Roles
| Role | How many | Hours / week | What they do |
|---|---|---|---|
| Data architect | 1 | 2-3 | Owns the spec and its decisions. |
| Analytics engineer | 1 | 2-3 | Tests the spec against a real model and questions. |
| Reviewer | 1 | 1 | Challenges claims and checks sources. |
9 · What you'll learn
- Why medallion layers and dimensional modelling are different axes.
- Why gold should be atomic and aggregates are a cache.
- The difference between a semantic compiler and a semantic runtime.
- Why dashboards should ship before chat.
Level-up questions
- How do you handle a question no certified metric covers?
- Who owns each metric definition as it drifts?
- When is a pre-aggregation worth its freshness cost?
10 · Links
- Links come once the work is public.