Skip to main content
You do not need to know the technical details — but having a rough mental model of what happens between “you hit send” and “Ana responds” makes a real difference in how you work with her.

​
The four-step process

When you send Ana a message, she works through four stages before responding. 1. Understand the question Ana reads your prompt and identifies the business concept (revenue, users, churn), the time frame, any filters you mentioned, and the format you seem to want. If your question is ambiguous, she will make a reasonable assumption and tell you what she assumed — or ask a clarifying question if the ambiguity is too significant to resolve on her own. 2. Select the right tools Ana picks the right combination of tools based on what your question requires. A question like “what were our top 10 customers last quarter?” goes straight to Text-to-SQL. A question like “build a cohort retention chart” goes to SQL first, then Python for the visualization. A question like “what are competitors charging?” goes to Web Search. 3. Retrieve and process data Ana writes and runs the query, processes the results, and checks whether they make sense. If a query returns zero rows or an unexpected shape, she will often try a different approach before surfacing the result to you. 4. Generate the output Ana formats the response based on what you asked for — a table, a chart, a written summary, or a combination — and explains what she did and any assumptions she made.

​
What Ana sees — and what she doesn’t

In scope for a given chat:
  • Connected connectors — the data warehouses or databases active when the chat starts. Ana can only query data from connectors that are selected for the session.
  • Context library — business rules, metric definitions, SQL examples, and documentation your admin has uploaded.
  • Ontology — pre-defined metrics and dimensions with consistent calculation logic.
  • Current thread history — everything said in the current conversation.
Out of scope:
  • Previous conversations — Ana has no memory across threads.
  • Other users’ threads — Ana does not see what your colleagues have asked.
  • Undefined business logic — if your company has a specific definition for a metric (e.g., “active user,” “churn,” “GMV”), Ana won’t know it unless it’s in the context library or ontology. Without a definition, she’ll make a reasonable guess that may not match your internal standard.
  • Data outside connected connectors — if a table isn’t in a connected warehouse, Ana can’t see it. CSV files uploaded directly into a chat only persist for that session.
Ana chooses her tools based on what the question requires, but you can also nudge her explicitly:
ToolWhen Ana uses itHow to nudge her
Text-to-SQLFetching data from your warehouse, filtering, aggregating, joining tables. Results are automatically loaded into the Python sandbox.”Query the database for…” or “Pull from the orders table…”
PythonCalculations beyond SQL, visualizations, statistical analysis, file exports”Build a chart showing…”, “Calculate the 30-day rolling average…”
Web SearchCurrent events, competitor research, external benchmarks — anything not in your data”Search for…”, “What does the market say about…”
Most analyses use Text-to-SQL and Python together: SQL retrieves the data, Python shapes and visualizes it. If you want a specific tool, just say so.

​
What affects answer quality

Prompt specificity. The more precisely you describe what you want — which table, which time range, which metric definition — the less Ana has to guess. See Writing Better Prompts for specific techniques. Context setup. If your organization has configured context (business rules, definitions, fiscal calendar), Ana uses it automatically. If that context is missing or outdated, Ana falls back to reasonable defaults that may not match your conventions. See What is Context to get started, or Context Scoping to control which context applies to which connectors. Data cleanliness and schema quality. Ana writes Text-to-SQL against your actual schema. Ambiguous column names, inconsistent naming conventions, or similar column names across tables are common sources of non-deterministic answers. The most effective fix is giving Ana explicit guidance through the context library — tell her which table is canonical for a given concept and how your schema is organized. For teams that want version-controlled, reviewable context, the GitHub integration is the most robust way to manage this over time. Connector state. Ana needs an active connector to query data. If Ana seems to be generating SQL but not returning data, check that the right connector is active. If it is active but still not returning data, the connector credential may have expired or the metadata may be stale — see Connector failing with a 403 or authentication error and Connector test passes but Ana cannot query the data in the FAQ. Older threads can also encounter errors if the underlying data has changed — starting a fresh chat is usually the fastest fix. See Chat Errors if you’re hitting a specific error code.

​
A note on threads

Ana treats each conversation thread as a self-contained session. She remembers everything within the current thread — earlier questions, filters you established, tables you referenced — but starts fresh in a new thread. This is a feature, not a limitation. It means you can start a clean analysis without prior context bleeding in. But it also means that if you want Ana to remember a definition or a filter you established earlier, you need to be in the same thread — or add it to your organization’s context library so it is always available. For guidance on when to stay in a thread vs. start a new one, see Writing Better Prompts.

​
Validating Ana’s Outputs

Every SQL query and Python operation is visible in the tool call cells — you can inspect exactly what Ana ran and verify it against your expectations. For exploratory analysis, Ana works well out of the box. For mission-critical outputs — board-level reporting, compliance-sensitive metrics — treat Ana’s outputs the same way you’d treat output from a new analyst: review the logic, validate against a known source, and build confidence incrementally. For a full guide on validation techniques, see Validating Ana’s Outputs.