Vendor Catalog

Divinci is a multi-vendor platform. Almost every step of the pipeline — embedding a chunk, storing a vector, parsing a PDF, searching the web, speaking a reply, fine-tuning a model, stripping PII — is a swappable tool, and most steps have several vendors behind them.

There are 64 such tools today, excluding the ~160 conversational models (those have their own live Model Catalog).

Why you would care

If you…	Look at
already pay a vendor and want to use that account	the BYOK column — attach your key and the tool becomes selectable
want to start with nothing	the Divinci-managed rows — usable with no key of your own
are choosing between two options	the Divinci usage column: what we actually run in production

The categories

RAG & Vector Stores Where chunks live and how they are retrieved — 14 stores, from Cloudflare Vectorize to Neo4j.

Embeddings The models that turn chunks into vectors, and why mixing them breaks search.

Parsing & Chunking Turning a PDF, a page or a record into chunks worth embedding.

Search & Scraping Web search providers and the scraping cascade.

Voice & Audio Text-to-speech, transcription, diarization, voice cloning and voiceprints.

Fine-tuning Where a fine-tune actually trains — Vertex, OpenAI, Cloudflare LoRA, Modal, Tinker, AWS.

De-identification Presidio and Stanford NER, and the one that fails closed.

Red-teaming & Evals First-party, and deliberately so — TrustBench probes, scorers and attested runs.

How vendor selection works

Every tool has an auth provider. A tool is selectable when credentials for its provider are available — either Divinci's (for the runtimes we pay for) or yours, attached as a Provider Key (BYOK).

Credentials are per workspace, and validated on save. Each provider declares a JSON-schema credential shape and a live validation call, so a bad key is rejected at the point you paste it rather than at the point a customer asks a question.

A vector's tool is fixed at creation

Vector store and embedding model are properties of the vector, not of the query. Re-pointing an existing vector at a different embedding model does not re-embed anything — and vectors from two models are not comparable even at the same dimensionality, so similarity search degrades quietly rather than failing. Create a new vector and re-ingest.

About the numbers on these pages

The Divinci usage column is generated, not asserted: a weekly job asks the production API for aggregate counts and commits them into these pages. It carries counts only — no workspace, release, customer or corpus is identified, and the endpoint that produces it is a $group-and-$sum with nothing else in the projection.

Two honest caveats about reading it:

• Zero does not mean bad. Several vendors are supported because a specific customer asked for them; the rest of the platform has no reason to use them.
• A count is a count of configurations, not of traffic. A vector store used by one enormous corpus and one used by two hundred small ones look similar here.

Last refreshed 2026-10-16.