Question 1

Is this BERTopic with a UI on top?

Accepted Answer

No. The pipeline is our own: seven stages — ingest, preprocess, embed, reduce, cluster, represent, name — each resolved from a component registry by config. You can swap the embedder, reducer or clusterer independently, which is the part an off-the-shelf pipeline does not give you.

Question 2

Why not just paste it into ChatGPT or Claude?

Accepted Answer

For twenty documents, do. The difference shows up at volume: a chat window gives you an impression of what it read, with no count behind it and no way to tell what it skipped. Dovre embeds and assigns every document in the run, so each topic carries an exact document count, a share of the corpus, a volume trend and the documents nearest its centre. It is the difference between a summary and a census.

Question 3

How large a corpus can it handle?

Accepted Answer

Ingest streams to Parquet in batches, so pulling in millions of rows never loads the source whole. A modelling run then works on up to 200,000 documents on the default settings — that is the `max_docs_in_memory` cap, and raising it trades RAM for corpus size. We would rather publish the real ceiling than a rounder one; it is already far past what fits in a chat context.

Question 4

Will I get the same topics if I run it again?

Accepted Answer

Yes, on the default profile. The hashing embedder is deterministic by construction, and the reducer and clusterer are seeded, so the same corpus and the same config produce the same topics. That matters if you intend to track a theme from one month to the next rather than re-interpreting a fresh answer each time.

Question 5

Do I need a GPU?

Accepted Answer

No — and you do not need to think about one. We run the infrastructure: the embedding service, the CUDA hosts behind it, the worker fleet and the storage. What that buys you concretely is the Quality profile — BGE-M3 embeddings, UMAP and HDBSCAN — at throughput, without anyone on your side provisioning a GPU box, sizing a batch or keeping a model server alive between runs. (For the curious: the Fast profile genuinely needs nothing but NumPy. That is an engineering fact about the engine, not a hardware bill we are passing to you.)

Question 6

So what am I actually operating?

Accepted Answer

Nothing. You point Dovre at a source, confirm which column is the text and which is the timestamp, and read the result. There is no cluster to run, no model to download, no pipeline YAML to own, and no MLOps hire between you and a topic map. The seven-stage engine is swappable by config because that is how we tune your runs — not a configuration burden we are handing over.

Question 7

Why is this possible now and not five years ago?

Accepted Answer

Two things changed. Multilingual embedding models got good enough that documents about the same thing land together whatever language they are written in — no translation step, no per-language pipeline, which is what made a mixed-language Southeast Asian inbox tractable at all. And every team now has an LLM that will happily read their text and give them an impression of it, which has made the absence of a count much more obvious than it used to be. Topic modelling is not new. Topic modelling that survives contact with five languages and 200,000 documents is.

Question 8

What does it cost?

Accepted Answer

We scope it per engagement rather than publishing a price list, because the two things that drive cost — how much text you have and where the run has to happen — vary more between customers than anything else. A single-tenant deployment inside your own network is a different proposition from a corpus we can run on our infrastructure. Tell us the corpus size and the constraint and we will come back with a number.

See it run on your own text.

We scope the question

We run it — all of it

You decide what happens next

Book a demo

Want the detail first?