Ollama & embeddings

The Ollama & embeddings section of Settings does two separate things that share a screen: it connects a local Ollama backend, and it picks the embedder your workspace uses. Embeddings turn text into vectors, which is what every semantic feature here runs on.

The two are independent. An embedder does not have to be an Ollama — it is a provider:model pair against any provider you have configured, and Ollama is one of the options rather than the assumption.

Local Ollama backend

Point Chatty at a running Ollama instance to use local models. Once configured, Ollama serves both chat/completion models and the embedding model.

A private address will not work

Chatty runs in Cloudflare’s network, so it cannot reach localhost, 192.168.x.x or anything else behind your router — typing that address into the form registers a server nobody can call. To use an Ollama on your own machine, run npx @chatty-lab/tunnel there instead: it connects outwards and registers itself here. See Connect your own Ollama.

The form below is for a server that is publicly reachable on its own.

The embedder

Each workspace has two embedder settings, resolved separately:

Setting	What it embeds
The system embedder	Context items, prompt search, the documentation index
The Vault embedder	The research corpus

Each is a provider:model pair — novita:baai/bge-m3, ollama:nomic-embed-text, nvidia:nvidia/nv-embedqa-e5-v5 — so the provider is part of the choice. Setting one and leaving the other empty is a real state and a common surprise: the docs index works and the Vault does not embed, in the same workspace on the same day.

1. Open Settings and go to Ollama & embeddings.

2. Pick an embedder for each of the two. If you are using an Ollama one, the model has to be pulled on that instance; for a hosted provider, the workspace needs its key.

3. If you changed a model, use Re-vectorize now so existing content is re-embedded.

A workspace with no credentials at all borrows one

The platform can lend its own embedder, but only to a workspace that has no keys and no Ollama of its own. A workspace with one key of its own pays for its own embedding — and if it has not chosen an embedder, it simply does not embed rather than falling back to a default. The screen says when what you are looking at is lent.

Changing an embedder means re-vectorising, promptly

Vectors from two models are not comparable, and the platform sends the task prefixes (search_query: / search_document:) that two-tower models expect, so anything embedded before that compares worse.

For context items it is sharper than “worse”: that search does not separate rows by model, so a new model with a different number of dimensions makes prompt search fail outright until the rewrite finishes. The Vault does keep models apart and simply re-embeds what it needs.

Re-vectorising a large workspace takes a while. Start it right after the change, not later.

What embeddings power

The system embedder drives:

• Semantic context — finding relevant context items by meaning for RAG.
• VectorDB Query nodes — similarity search in graphs; see the VectorDB Query node.
• Episodic agent memory — how an agent recalls relevant past turns; see Memory strategies. Each episode is stored with the model that embedded it, so changing the embedder here restarts recall for conversations already under way. (Semantic memory does not use embeddings — it extracts facts with a chat model.)

Memory strategies How agents remember across turns, including embedding-backed memory.