AI & context
7 min

Questions to Ask Before Buying a Knowledge Graph Platform

Key takeaways

  • Ask for demonstrations instead of confirmations. A product page can answer any yes-or-no question.
  • Exit questions first: export the model, read the file, then re-import it elsewhere.
  • The most revealing question is rarely asked: change a definition, then show me what it said before.
  • An evaluation missing the future owner and the consuming-system engineer optimises for the purchase.
In this article
Share

What should you ask before buying a knowledge graph platform?

Questions phrased so the answer cannot be a feature name.

Vendors in this market answer honestly in my experience. The problem is almost never that somebody misled a buyer; it is that the buyer asked questions a product page could answer. "Do you support SKOS" gets a yes from everyone. "Export this model and show me the file" gets a different conversation.

What follows is organised by who should ask it, because the questions that catch real problems come from people who are often absent from the evaluation.

What should you ask about getting out?

Ask these first, because they are the hardest to fix later and the least likely to be volunteered.

Exit questions

  • "Export this model with its constraints and show me the file." Read what actually came out. Constraints and provenance are the usual casualties.
  • "Can I re-import that into a different product?" The step nearly everyone skips, and the only one that proves portability.
  • "Does the export include who approved each definition and when?" Provenance is the hardest layer to recreate and the easiest to lose.
  • "Which parts of my model live in your proprietary extensions?" Most products have some. The answer should be specific.

Open standards vs proprietary graph platforms covers why these four decide more than anything else on the list.

What should you ask about governance?

These separate products that look identical in a demo.

Governance questions

  • "Change this definition, then show me what it said before." Versioning, audit logs and approvals often exist as three features with nothing joining them.
  • "Who approved this, by name?" A last-modified stamp is a different thing from an approval.
  • "How long does an approval take in a real deployment?" Latency is a governance property, and a slow loop means people route around the vocabulary.
  • "Can a domain expert change a constraint without a deployment?" If not, every definitional change queues behind engineering.

The first question is the single most revealing one on this page, and it is rarely asked. Reconstructing a past definition requires the version, the approval and the content to be joined, and many products store all three and join none of them.

What should you ask about consumption?

Ask these with whoever runs the system that will actually call the platform, because nobody else will raise them.

Consumption questions

  • "Resolve a term from curl, and time it." If the answer involves a connector or an export, retrieval systems will not use it.
  • "What happens when a term is unresolved?" Silence gets treated as absence of ambiguity, which is worse than an error.
  • "Can a consumer pin to an agreed version?" Downstream systems need to keep reading one version while the next is drafted.
  • "Show me the rate limits." A governance API called on every query has a very different profile from one called nightly.

What should you ask about the commercial shape?

Three, and the second is the one that produces surprises in year two.

"What is in the enterprise tier?" Governance features sitting above the tier you are buying is a common and legitimate packaging choice. It is also a cost you should see now.

"How does pricing move as consumers grow?" A platform priced per seat behaves very differently from one priced per call once AI systems start resolving terms at volume.

"What does support actually cover?" Specifically whether modelling help is included or billed, because the first year is mostly modelling.

What do good and bad answers sound like?

The pattern is consistent enough to be useful. Good answers are specific and often include a limitation.

You askA good answerA warning sign
Export with constraintsProduces a file, names what is omitted"Fully supported" with no file
What did this say beforeShows the prior version with its approverOpens an audit log of events
Resolve from curlRuns it, shows the latencyDescribes an integration pattern
Approval latencyQuotes a real customer's numberDescribes the workflow designer

A vendor who volunteers a limitation usually earns more trust than one who does not. Every product in this category has limitations, and the hidden ones are the ones you meet later.

Who needs to be in the room?

Four people, and two of them are usually missing.

Required attendees

  • A domain expert who will approve definitions. They judge the editing experience in ten minutes and nobody else can.
  • Whoever runs the consuming system. The consumption questions come from them or from nobody.
  • Someone who has migrated a platform before. They ask the export questions the room finds tedious.
  • The person who will own this in year three. Often unnamed, and that absence is itself a finding.

An evaluation without the last two optimises for the purchase instead of the operation, and the product that wins is the one that demos best. How to evaluate a knowledge graph platform covers the surrounding process.

When in the process should you ask these?

Earlier than feels polite, and the exit questions belong in the first conversation.

The usual sequence puts portability last, somewhere near contract negotiation, by which point the team has invested weeks in a preferred product and nobody wants the answer to be awkward. Asking on the first call costs nothing and changes which products reach the trial.

A workable order

  • First call: exit. Export, re-import, provenance. Two of your shortlist will quietly drop out here.
  • Trial week one: governance. Change a definition and ask for the previous one, on your own model.
  • Trial week two: consumption. With the engineer who will call the API, timing real requests.
  • Before contract: commercial. Tier boundaries and how pricing moves as consumers multiply.

Teams that run it in that order finish faster, because the expensive discoveries happen while the shortlist is still cheap to change.

So what should I take away?

Ask for demonstrations instead of confirmations. Export a model and read the file. Change a definition and ask what it said before. Resolve a term from curl and time it. Every one of those produces an answer a feature list cannot.

TopQuadrant is one of the vendors you would be asking, so apply all of this to us. Our knowledge graph platform is built on RDF, OWL and SHACL specifically so the export question has a good answer, and the right response to that sentence is to test it. If we fail one of these, finding out during a trial serves you better than finding out in year three.

Give your AI the context it's been missing

See how the TQ Data Foundation turns your enterprise knowledge into trusted, Al-ready context.