Can I compare them on my own corpus?

Yes, and this is the honest way to decide. The console can take any search from your history and re-run it against RAG-only retrieval, then show you both answers with both traces. The comparison run is logged underneath the original with the reason attached, so you build up a record rather than a vague impression.

Is the graph tier slower?

Yes. It adds a model call to extract entities from the question, a two-hop traversal, and a second retrieval pass for the evidence lane. The direct lane still finishes at RAG speed — fusion just waits for the other two. Every response carries per-step timings so you can see the real number for your corpus rather than trusting ours.

What happens to documents I ingested before upgrading?

They stay indexed and searchable, and they are not retroactively extracted into the graph. If you want them in the graph, re-send them — ingest is an idempotent upsert on your own identifier, so a full re-POST of your corpus is safe and replaces rather than duplicates.

Can I go back to RAG?

Not on the same workspace; the upgrade is one-way. Create a second workspace on the free tier if you want a permanent side-by-side rather than a per-query comparison.

Comparison

When is a knowledge graph worth the extra cost?

Sometimes it is not, and this page says when. Both tiers are the same two calls over the same content — the difference is how many ways a question gets asked, and what the answer can say as a result.

See the retrieval engine Start free

At a glance

One lane, or three.

RAG tier compared with the graph tier
	RAG tierfree	Graph tier
Retrieval
Retrieval lanes	1 — direct hybrid	3 — direct, expanded, evidence
Reaches content the query never worded	Not included	Included
Evidence restricted to provenance documents	Not included	Included
Cited answers with enforced grounding	Included	Included
Per-step diagnostics	Included	Included
What ingest returns
Response	200 · indexed	202 · extracting + statusUrl
Searchable immediately	Included	Included
Graph populated	Not included	asynchronously
What the answer can say
Summarise a passage	Included	Included
Combine two passages the query matched	Included	Included
State a relationship no passage contains	Not included	Included
Name every document behind a claim	Not included	Included
Governance
Versioned ontology and drift	Not included	Included
Graph visualizer, entity quality	Not included	Included
Corpus health	Not included	Included

RAG tier

free

Retrieval lanes: 1 — direct hybrid
Reaches content the query never worded: Not included
Evidence restricted to provenance documents: Not included
Cited answers with enforced grounding: Included
Per-step diagnostics: Included
Response: 200 · indexed
Searchable immediately: Included
Graph populated: Not included
Summarise a passage: Included
Combine two passages the query matched: Included
State a relationship no passage contains: Not included
Name every document behind a claim: Not included
Versioned ontology and drift: Not included
Graph visualizer, entity quality: Not included
Corpus health: Not included

Graph tier

Retrieval lanes: 3 — direct, expanded, evidence
Reaches content the query never worded: Included
Evidence restricted to provenance documents: Included
Cited answers with enforced grounding: Included
Per-step diagnostics: Included
Response: 202 · extracting + statusUrl
Searchable immediately: Included
Graph populated: asynchronously
Summarise a passage: Included
Combine two passages the query matched: Included
State a relationship no passage contains: Included
Name every document behind a claim: Included
Versioned ontology and drift: Included
Graph visualizer, entity quality: Included
Corpus health: Included

The difference, concretely

What a graph answer can say that a chunk cannot.

Both tiers find both of these passages. Only one of them can state what they add up to.

The question

Does finishing onboarding early actually reduce churn?

Two documents each hold half of it. Neither one contains the answer.

Cohorts that completed onboarding within seven days retained at 91% by day ninety, against 78% for cohorts that took longer.
Q3 Retention Review

The strongest predictor of a retained account is reaching first value quickly. Accounts that stall before activation raise tickets about scope, not about the product.
Support Themes 2026

The fact the graph holds

Onboarding completion —reduces→ Ninety-day churn (confidence 0.95, 3 sources)

The answer

Cohorts that complete onboarding within seven days show lower churn at ninety days [1], and the support record attributes the difference to reaching activation rather than to onboarding itself [3].

The RAG answer is not wrong — it will summarise both passages accurately. It just leaves the reader to notice that the second passage explains the first, because nothing in the corpus says so out loud. This example joins two documents; the one on the home page joins a billing record to a help article, which is the same mechanic across two applications rather than two files.

What it costs

The upgrade is not free, in either sense.

Three real costs, worth knowing before you decide.

What you give up

What you get for it

Latency. Entity extraction on the query, a two-hop traversal, and a second retrieval pass all happen before fusion.

The expanded and evidence lanes, which is what reaches content the question never worded.

Money. A model call per shard per document at ingest. Roughly a fifth of a cent per thousand documents, but unbounded if you point a firehose at it — which is why the tier needs a payment method.

A graph that persists, is queryable, and is rebuildable without paying that cost again.

Your existing corpus is not retroactively extracted. Upgrading changes what happens next, not what happened.

Re-POSTing is safe and idempotent, so backfilling is a loop rather than a migration.

Honest fit

When plain RAG is the right answer.

Most vendor comparison pages do not have this section. It is the most useful one here.

Your questions are answered by single passages
If people ask "what is our refund window" and one document says, the graph adds latency and cost for nothing. Lookup is not the problem a graph solves.
Your content is one system, with one vocabulary
The graph earns its keep by joining entities across applications — the account in your billing system and the account in your ticket queue. With one source there is nothing to join, and cross-source links, the metric that tells you it is working, will read zero.
You need the lowest possible latency
A typeahead or an in-product search box that has to feel instant is a direct-lane job. Run RAG there and use the graph tier for the analytical surface where a second of thinking is acceptable.
You are still finding out whether retrieval helps at all
Start free, ingest a few thousand documents, see whether the answers are useful. If they are not, the graph will not rescue it — the problem is upstream, in the content.

We cannot show you an accuracy number here.

Graph fusion contributes to 201 of 231 question runs on our evaluation corpus, so the extra lanes are doing real work. But on the question sets we can score, the graph has not been shown to answer a question that RAG-only retrieval missed.

So the case for upgrading is the graph itself — the cross-document relationship, the evidence trail, the governable vocabulary — not a percentage point of accuracy. If accuracy is the only thing you are buying, stay on the free tier and tell us; that is useful information for us and a saved invoice for you.

Read the full methodology

FAQ

Frequently asked

Compare them on your own content.

Start on the free tier, then run any query against both from the console when you want to see the difference.

no card required
first workspace is free
one POST to ingest

Talk to us Start free