Business meaning before business answers

1. Why a correct query can still produce a wrong answer

Imagine a product manager asking, “How did weekly active users change?” A database can count rows perfectly while answering the wrong business question. Does “active” mean opening a page, submitting a question, or completing an analysis? Are mobile users included? Are internal test accounts excluded? Is one person using three surfaces counted once or three times?

These are not SQL syntax questions. They are questions about business semantics: the meanings, rules, and relationships an organization has agreed to use. An agent that guesses them can produce a polished report with misleading numbers.

The article describes Genie Ontology as a continuously updated understanding of institutional knowledge. It begins with human-curated, governed concepts in Unity Catalog, then supplements them with knowledge learned from dashboards, documents, queries, notebooks, and connected applications.

Here, “ontology” should be read in the product’s practical sense: business definitions, rules, relationships, and trusted sources. The post does not specify a formal ontology language, an inference calculus, or a particular graph-storage design.

Four terms worth separating

2. The mechanism: choose meaning, retrieve evidence, compute

The demonstrated task is a weekly review of Genie One adoption. Users interact across web, desktop, mobile, Slack, and Teams. Some dashboards are authoritative; others are stale. Plans are in Google Docs, bugs in Jira, and feedback in Slack. The report must combine these different kinds of information rather than treat them as interchangeable text.

Conceptual reconstruction of the demonstrated workflow, not a disclosed internal architecture.

Why supplying context can reduce exploration

The post contrasts this workflow with a general-purpose agent that may crawl schemas, sample tables, read documents, and run repeated queries to discover the right definitions. Pre-existing business context can narrow that search before and during reasoning. This is a plausible mechanism for reducing effort and misinterpretation, but the article provides no measured latency or token-cost comparison.

3. From one answer to an ongoing workflow

The article continues beyond the initial report. These later steps matter because the ontology supports a recurring product-management activity, not just a one-off lookup.

The summary also mentions custom skills and MCP writes to external tools. Those are stated capabilities, not mechanisms evaluated in detail in this walkthrough.

4. Playground: popular is not the same as authoritative

Two assets both match the words “weekly active users.” One defines a governed metric across surfaces; the other is a popular legacy dashboard that counts web page-open events. Change the ranking policy and see how the apparent adoption trend changes.

This is an educational simulation, not Genie, OntoRank, or a recreation of Databricks data. All records, scores, definitions, and permissions below are invented.

Three experiments to try

1. Move certification weight to zero. The popular legacy asset wins. Its exact match to the words does not make its event-count definition match the governed KPI.
2. Reset, then disable the mobile feed. Even the right definition undercounts when evidence is incomplete. Authority is not data completeness.
3. Change the user role. Restricted explanatory evidence becomes visible only to the authorized role. A higher source score could not make it visible to the product manager.

5. Work through the example by hand

In the complete synthetic feed, Week A has questions from external users u1, u2, u3, u4. Week B has questions from u1, u2, u3, u4, u5, u6. A repeated question by u1 is still one active user. An internal account is excluded. The governed result is therefore 4 → 6 users, or 50% growth.

The legacy dashboard instead counts web page-open events: two in Week A and two in Week B. It reports 2 → 2 events, or 0% growth. This is not a computational error. It answers a different question.

With the mobile feed unavailable, the governed query sees only three active users in each week, producing 0% observed growth. The definition remains correct, but the evidence is incomplete. The report should flag that condition rather than present the result as complete cross-surface adoption.

6. A small, inspectable prototype

The playground uses a deliberately simple ranking formula:

score = w × certification
      + (1 − w) × (0.7 × relevance + 0.2 × usage + 0.1 × authorship)

This formula is invented for teaching. The article names certification, usage, and authorship as authority signals, but does not disclose OntoRank’s algorithm or its weights. Here, the invented relevance feature makes the contrast easy to explore.

A minimal context-assisted reporting loop can be written as follows. The stricter compatibility check below is an implementation recommendation; the playground intentionally allows a mismatched candidate to show the failure mode.

// All policies and schemas below are illustrative.
function review(user, request, sources, events) {
  const visible = sources.filter(s => mayRead(user, s));
  const compatible = visible.filter(s =>
    s.concept === request.concept &&
    s.definitionVersion === request.definitionVersion);
  if (!compatible.length) return {status: 'needs clarification'};

  const source = compatible.sort((a,b) => score(b)-score(a))[0];
  const data = events.filter(e => mayRead(user, e));
  const counts = computeUsing(source.definition, data);
  const notes = searchLiveTools(user, request); // access enforced here too
  return {
    counts, coverage: checkCoverage(data, request),
    citations: [source.provenance, ...notes.map(n => n.provenance)],
    explanations: notes, // associations, not automatic causal claims
    definitionVersion: source.definitionVersion
  };
}

The essential separation is between selection and execution. Context chooses the rule and the source; deterministic analysis computes the measure. The answer carries the definition and provenance so a reviewer can inspect both.

For a production system, access control belongs at retrieval, tool access, and query execution—not just in this illustrative client-side filter. Learned snippets also need source links and handling for outdated or conflicting definitions. The blog describes automatic maintenance, but does not explain those maintenance procedures.

7. What the article establishes—and what remains untested

This is a product-use walkthrough, not a controlled evaluation. Its evidence is a narrated internal workflow with screenshots and interface examples. Its reported outcome is an executive-ready review, follow-up investigation and forecasting, recurring scheduling, a follow-up ticket, and a shareable agent.

Practical failure modes to keep in mind

Authority can lag reality. Certification is valuable, but a formerly correct definition may become outdated. High usage can reflect familiarity rather than truth. Authorship is a trust signal, not a universal endorsement. These are general risks of the design pattern, not failures measured in the article.

Permission-aware answers may be partial. An agent should not disclose restricted context, but it should avoid turning missing evidence into a confident explanation. A shared agent cannot simply inherit the creator’s effective access for everyone.

Citations make checking possible, not automatic. A citation can identify a source without demonstrating that the claim follows from it. For a KPI, the reviewer still needs the chosen definition, filters, time window, and data coverage. For a forecast, the reviewer needs assumptions and uncertainty.

Recurring workflows amplify both good and bad choices. Scheduling saves repeated effort, but repeated semantic mistakes also scale. A robust review process would monitor definition changes and missing feeds and require appropriate approval for external writes. The post does not describe these controls in detail.

8. One closing connection

For a context platform, the transferable lesson is to preserve the distinction between governed meaning, learned context, and live evidence—and attach permissions and provenance to each. The article supports that pattern, but does not provide a blueprint for a context lake or a specific context-graph implementation.