1
Opens the add modal — the provider is chosen from a grouped preset catalog. Once added, its model catalog (below) fills in automatically.
2
The status of the last connection test. “Test” raises a
toast
with the verdict.
3
Provider menu: edit · test · delete. “Delete” is blocked while any of its models is assigned to a function or is in chat (assignment by function).
4
The API key is shown only as a mask; local providers usually have none.
5
When a provider errors, its models are unavailable: the affected
functions wait for the connection to recover or for the Admin to reassign. The rest of the platform keeps running.
6Platform (is_system) — the platform's built-in embedder, out of the box and with no external provider. Can't be deleted, has no key.
7Provider is picked from a grouped preset catalog (cloud vendors · aggregators · Chinese clouds · self-hosted). The preset carries both kind (Cloud / Self-hosted) and adapter, and adapts the form: a cloud vendor requires a key; a self-hosted runtime requires a base URL and leaves the key optional.
8Base URL is prefilled by the preset (Chinese clouds, aggregators, self-hosted) and stays editable. Native-SDK clouds (OpenAI / Anthropic / Google) hide the field — the SDK knows the endpoint.
9Self-hosted — the company's own runtime (vLLM · TGI · Ollama). Most speak the OpenAI dialect, so they ride the same openai_compatible adapter as OpenAI-compatible clouds; kind — not the adapter — is what separates own hardware from a cloud.
Provider catalog and activation
The selected provider's models: discovery fills the catalog automatically, the rest the Admin adds by hand. From the catalog a model is enabled on the platform right in the row.
1
Provider selector — the table below it redraws for that provider's models.
2
Opens the manual-entry modal — for providers without discovery and for fine-tuned / custom models (source “manual”).
3
The model's display name in assignment and spend. The pencil edits it right in the row.
4
The type (chat / embedding) decides which functions the model can be assigned to. Edited in the row; the change is blocked while the model is assigned to a function or is in chat.
5
“Enable” activates the model, “Disable” returns it to the not-enabled set. Disabling is blocked while the model is assigned to a function or is in chat — reassign first.
6
An embedding model carries its vector width. Only a model whose width matches the knowledge base can be assigned (assignment by function), so it's shown right on the row.
7
Embedding dimensions — the model's vector width, entered here because discovery doesn't report it. Built-in models come pre-declared; the field appears only for the embedding type.
Every platform function consumes a model.
System functions (embedding) — one model per function, assigned by the Admin. Chat and agents — a default model plus a list the employee or the agent's owner picks from.
platform functions
System functions
Function
Model
Embedding
AI
Cohere embed-v3 ▾
Cohere embed-v3 Cohere · 1024 · currentbge-m3 Platform · 1024Qwen3-Embedding-0.6B Platform · 1024Embedding 3 Small OpenAI · 1536 · needs 1024Embedding 3 Large OpenAI · 3072 · needs 1024Nomic Embed Ollama · dimension not set
1
assigned
User chat
AI
The employee picks the chat model themselves — from this list;
the primary one opens by default. 2
bge-m3 Platform · 1024 · out of the boxQwen3-Embedding-0.6B Platform · 1024 · out of the boxCohere embed-v3 Cohere · 1024Embedding 3 Small OpenAI · 1536 · needs 1024Nomic Embed Ollama · dimension not set
The platform will regenerate every vector under the new model —
a bulk re-embedding (≈13.1k chunks). A background run will
start; while it's in progress, changing again is unavailable.
Cohere embed-v3→bge-m3
Search keeps working during the recompute; the run's
status is in
Knowledge Store.
1 Embedding — one model for indexing and search; the assignment is unified, or search breaks. There's no default: the platform option is available out of the box, but embedding waits for the Admin's explicit choice (built-in or your own). Assigning warms the runtime ahead of time — the first search doesn't wait on the weights. The knowledge base stores fixed-width vectors — 1024 in v1 — so only models of that dimensionality can be picked; the rest stay listed but disabled, with the reason inline. A cloud model doesn't report its dimensionality on discovery, so it's declared when the model is added to the catalog. Changing the assigned model is a special case: the new vectors are incompatible with the old, so the choice opens a confirmation that starts a full re-embedding. While the run is in progress, the row shows its progress and changing again is blocked (refresh).
2 User chat — the employee's choice: the Admin assembles the list of allowed chat models and marks one as the default. That same chat model runs the conversation and decides for itself whether to reach into the knowledge base.
3 Agents — a separate list of chat models with one default; the owner picks the agent's model from it. Removing a model here stops the agents that chose it, until they're switched to another.
Agent Engine.