Ask AI

One conversational assistant in the Explorer that answers questions about your context graph and helps you enrich, clean, deduplicate, and reshape your data — all from the same chat.

Where to find it

Open the Explorer and look for the ✦ Ask AIpanel on the right. Pick a context graph (and optionally a type) so the assistant knows what you're working with, then type a request in the box at the bottom. When the panel is empty it offers a few suggested questions and one-click actions to get you started.

How a conversation works

You type a request and the assistant replies with one of three things, depending on what you asked:

An answer

For questions, you get a written answer and — when there are results — a table. Expand View SPARQL to see the exact query, and copy it if you want to reuse it.

A clarifying question

If your request is ambiguous, the assistant asks for the missing detail instead of guessing — for example which type to enrich, or which field counts as a duplicate. Just reply in the chat and it continues.

A proposed plan

For anything that changes your data (enrich, clean, deduplicate, ontology edits), the assistant proposes a plan before it touches anything. Each step shows what it will do, a short rationale, a confidence level, and an estimated cost — including a clear Paid badge and dollar estimate when a step uses paid web lookups. Nothing runs until you press Confirm & run (or Cancel).

When you confirm, the work runs in the background as one or more jobs. The chat shows a “Done” summary with a View in Jobs link where you can watch progress and review results. Multi-step plans — like cleaning a field and then enriching it — are sequenced automatically, in the right order.

A quick example

A typical clean-then-enrich exchange looks like this:

text

You    ▸ Fill in the website for every Company

Ask AI ▸ Proposed plan · 2 steps

         1  clean   Normalize Company.name
            Names contain emojis and trailing punctuation that
            hurt web-lookup accuracy.
            98% confident · No paid calls

         2  enrich  Look up Company.website
            312 companies are missing a website.
            87% confident · 312 paid calls · ~$1.56

         [ Confirm & run ]   [ Cancel ]

You    ▸ (clicks Confirm & run)

Ask AI ▸ Done
         • clean: normalized 312 names
         • enrich: queued for 312 companies — job enr_xxxxxxxx
         [ View in Jobs ]

The assistant noticed that cleaning the names first would make the website lookup more reliable, so it added that step on its own — and showed you the cost of the paid step before anything ran.

What you can ask

It's all the same chat — you don't pick a mode. Describe what you want in plain language and the assistant figures out which capability fits. Here are the five things it can do, with prompts you can paste in to try.

Answer questions about your data

Ask anything about a context graph in plain English. The agent reads your ontology, runs the query for you, and replies with a written answer plus a results table. You can expand “View SPARQL” to see exactly how it got there.

Try asking

How many companies are in this graph?
Which suppliers ship to more than three regions?
Show me people with no email on file
What’s the average deal size by region?

Enrich attributes from the web

Have sparse columns? Point the agent at a type and the attributes you want filled and it looks the values up from external sources — a company’s website, an industry, a headquarters. Before anything runs you see how many entities are affected and an estimated cost, so paid lookups never happen by surprise.

Try asking

Find the official website for every Company
Fill in the industry and headquarters for suppliers in the EU
Enrich Person.employer where it’s missing

Clean and normalize data

Messy values get in the way of good answers. The agent can split a multi-value field into separate atomic entities, strip emojis and stray punctuation, and standardize formatting. When cleaning would make a later step work better, it proposes the cleanup first — for example, tidying names before it tries to enrich them.

Try asking

Split the skills column into separate Skill entities
Strip emojis from every name and title
Normalize the country field to consistent values

Deduplicate entities

Real data has duplicates — the same company entered three ways, two records for one person. Ask the agent to find likely duplicates with fuzzy matching and merge them, so each real-world thing is a single entity in your graph.

Try asking

Find and merge duplicate Company entities
Are there duplicate people in this graph?
Deduplicate suppliers using fuzzy matching

Inspect or extend the ontology

The ontology is the shape of your graph — its types, their attributes, and how they relate. Ask the agent to walk you through a type, or to add a new attribute or a whole new type when your data grows beyond its original schema.

Try asking

What attributes and relationships does Company have?
Add a ‘founded_year’ attribute to Company
Create a new type called Contract linked to Supplier

Tips

Select a type first. Choosing a type in the Explorer gives the assistant context, so you can say “enrich the website” instead of spelling out the full type name every time.
Read the plan before confirming. The estimated cost and affected-entity count are right there — check them, especially when a step is marked Paid.
You can keep talking. Each chat keeps context across turns, so you can refine a request, answer a clarifying question, or follow up on a previous answer. Press New to start a fresh conversation.
Track work in Jobs. Anything that runs in the background shows up on the Jobs page, where you can monitor progress and review what changed.

Next steps

Quickstart — get data into a context graph so there's something to ask about
MCP Server — use the same capabilities from Claude, Cursor, or any MCP-compatible agent