Skip to content
ViziGraph Wiki

Search is only available in production builds. Try building and previewing the site to test it out locally.

2-how_it_works.md

How the AI Assistant Works

The assistant answers questions about your data by exploring the knowledge graph for you.

It Reasons in Steps

Rather than translating your question into one query, the assistant runs a short sequence of steps, choosing each one from what it just learned:

  1. Read the ontology — which object types and relations are relevant
  2. Ground the names you used — find the matching real objects
  3. Plan a route — how those objects are connected
  4. Query the graph and present the result

Progress messages show you which step it is on.

What You Get Back

  • A table — when the answer is a list of records
  • A chart — when the data suits a visualization
  • Text — for a single value, a yes/no, or an explanation
  • A question — when your request is genuinely ambiguous

Tables and charts come with a written interpretation, not instead of one.

When It Asks You Something

Questions arrive as a short inline form (up to four questions, with clickable options) that replaces the message box — never as a vague “could you clarify?” paragraph. You will see one when:

  • Your request has no clear scope (“analyse this”)
  • Several records match the name you gave
  • A complex task needs its plan confirmed before it runs

Pick an option, or use Other to type your own. The request resumes where it paused. If you leave it, the assistant proceeds on its best assumption after a while and tells you which one it took.

It Remembers the Conversation

Within a conversation, the assistant keeps what you established — the objects you named, the values it confirmed. Follow-ups work naturally:

“How many in total?” — “And for the other one?” — “Show that as a chart.”

Start a new conversation (+) when you change topic, so old assumptions are not carried over.

Getting Better Answers

  • Ask in full sentences — the assistant reads intent, not keywords
  • Use exact identifiers when you know them
  • Ask one thing at a time
  • Refine (“only the active ones”) instead of re-asking from scratch
  • Correct it in your next message if it misunderstands

Languages

Your question is answered in the language you asked it. Supported: English, French, Spanish, Italian, German.

Limits

Each question gets up to 12 reasoning steps. A very broad question can reach that limit — the assistant then still answers with what it found and tells you what it was doing, rather than failing silently. Narrowing the question usually resolves it.

Large results are capped for display, and the assistant reports the true total when it truncates. Long queries time out after a few minutes; add a filter to reduce the search.

Reopening an old conversation shows tables and charts as empty shells — click to load the data, which is always fetched live.

What It Will Not Do

  • Never modifies your data — it is read-only
  • Never invents links or names — if something is absent, it says so
  • Returns no personal data
  • Reports zero as a real answer rather than loosening your question until something matches

Feedback

Each answer has 👍 / 👎 and a View details link showing what the assistant actually did. A 👎 on a wrong answer is the most useful signal you can give.