Model catalog
The platform ships four independent catalogs: chat models (LLM), embedding models,
rerankers and media models. Each catalog is a set of YAML files on the server
(config/llm, config/embeddings, config/rerank, config/media). One file describes
one provider — its class, base URL and the environment variable holding its key — plus a
list of models.
Model ids
A model is addressed by a compound id, <provider>/<model>, where the first segment is
the provider name from the YAML and the rest is the provider’s own raw id:
| Example id | Catalog | What it is |
|---|---|---|
| openrouter/openai/gpt-4o-mini | llm | an OpenAI model served through OpenRouter |
| sentence_transformers/all-MiniLM-L6-v2 | embeddings | local embeddings, no key needed |
| openrouter/cohere/rerank-v3.5 | rerank | a reranker through OpenRouter |
| ws:9f2c…:gpt-4o | any | a model from your own connection (BYO) |
There can be two slashes: the split happens on the first one, so
openrouter/cohere/rerank-v3.5 correctly means provider openrouter, model
cohere/rerank-v3.5.
What is in the catalogs
- LLM — the OpenRouter catalog (several hundred models, regenerated by a script from the vendor’s live API) plus a catalog for a local OpenAI-compatible server if your administrator configured one. Every entry carries its tariff in credits.
- Embeddings — local
sentence-transformers(they need no key), OpenAI and OpenRouter. For many models the vector size is known in advance and pinned in the file. - Rerankers — OpenRouter, Cohere, Jina, Voyage. They all speak the same Cohere-compatible protocol, so a new vendor is a YAML file rather than new code.
- Media — speech-to-text, text-to-speech and image generation. Each entry carries its modality and a schema of its own parameters.
The entry marked as the catalog default is lifted to the top of the list — that is the model a freshly dropped node starts with.
Where a model is chosen
| What you configure | Where |
|---|---|
| Chat model | the Model field of the llm node — a config node wired with a link_llm edge into llm_response, ai_agent or structured_output_llm |
| Embeddings | the embedding node, and the model picker when you create a knowledge base |
| Reranker | the rerank node |
| Media | image_generate, text_to_speech, speech_to_text |
| Web search | the openrouter_web_search node — OpenRouter models only |
Media nodes render a parameter form that depends on the chosen model: the platform serves the editor that model’s own schema, so switching models changes the fields (a voice for speech synthesis, size and quality for an image).
Visibility and price
Settings → Models decides which models appear in the LLM node dropdowns. Models are grouped by source — system, workspace and chains — and each row shows its price in credits per 1000 tokens, input and output separately.
Models
System models
A disabled model disappears from the dropdowns
| Model | Price per 1000 tokens | Visible |
|---|---|---|
| openrouter/openai/gpt-4o-mini | 0.03 / 0.12 cr | on |
| openrouter/qwen/qwen3-8b | 0.024 / 0.091 cr | on |
| openrouter/openai/gpt-oss-20b:free | 0 / 0 cr | off |
The prices above illustrate the format, not a price list: the current tariff of each model is shown in the section itself. A model that is not in the catalog is billed at the default tariff — it is never treated as free.
Embeddings and knowledge bases
A knowledge base pins its vector size once, at creation: either from the catalog, or by probing the endpoint with one call. From then on the stored value wins over the catalog — otherwise editing a YAML file would turn a working base into a pile of incompatible vectors. So you cannot change the embedding model of an existing base; create a new one and re-index the documents.