Providers

Admin → Providers does two related jobs: it holds your LLM connections, and it configures which of them SEMANTIX uses itself. It is the first thing to set up — without a provider, Chat does not work.

Adding a provider

• Name — unique. A duplicate name is rejected.
• Base URL — an OpenAI-compatible endpoint, e.g. https://openrouter.ai/api/v1.
• API key — stored encrypted. It is never displayed again, only masked.
• Model — the model id the endpoint expects.
• Context window — default 32768. This feeds token budgeting, so setting it accurately matters: too high and the agent tries to fit more context than the model accepts.

Defaults

• Set as default — the provider Chat uses. Exactly one provider can be the default; setting a new one clears the previous automatically, so there is never a window with two.
• Set as DSPy default — the provider used by the metric-generation pipeline. Also single. Independent of the Chat default, so metric generation and everyday chat can use different models.

Your own gateway

base_url is the whole integration point. Pointing it at your organisation’s approved AI gateway means you use the models your policies allow, and prompts and data stay inside your boundary. Any OpenAI-compatible endpoint works.

The SEMANTIX models

Separate from the provider list, this section picks which models SEMANTIX uses for its own work:

• the generation model used for metric generation and related pipelines;
• the embedding model used to index entities and metrics for semantic search.

Reindexing embeddings

Change the embedding model and everything indexed with the old one is incomparable with the new one. Reindex embeddings rebuilds the index against the current model.

Caution

Reindex after changing the embedding model, before expecting semantic search to work. Until you do, results are effectively arbitrary — old vectors and new ones are being compared as though they meant the same thing.

System providers

One provider and the Local Embedder are marked as system providers and cannot be edited or deleted. This is deliberate: they are what the app falls back to, and removing them would leave a broken installation rather than a simpler one.

Troubleshooting

Chat returns Missing Authentication header or a 401. The stored key is not valid for that endpoint — almost always a pasted password, or a key from a different provider. Check that the key belongs to the base URL you configured.

Chat returns a model-not-found error. The model id does not match what the endpoint serves. Model names are provider-specific.

Nothing generates. No provider is marked as the default.