You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Set AI providers and recommendations from deployment configuration
#4653
Silo can take server settings from the environment (#4498) and install plugins from a provisioning file (#4602). AI providers and recommendations are in neither, so a deployment kept in git still needs a person in the admin UI, or a script.
Turning on recommendations means four manual steps: connect an embedding provider (type, base URL, key, what it serves), choose its model, enable recommendations, restart. The provider key already comes from a secrets manager in my setup, and the rest is configuration that belongs next to the compose file.
PR #4498 says that variants reading a secret from a file were left out of that change. A provider key is exactly that kind of setting, and the provider registry from #3106 has no way in from outside the admin pages.
Who it affects
Operators who deploy Silo from Compose, Kubernetes or NixOS and keep the rest of the server in version control. The same case as #4594 for transcode node keys.
Current workaround
Script the admin API after first start: POST /api/v2/admin/ai/providers (with /check and /{id}/models), then PUT /api/v2/admin/settings/recommendations.embedding_provider_id, ...embedding_model and ...enabled, then restart. This works but needs an admin API key as a deployment secret, can only run once the owner account exists, and reports failure in logs rather than at deploy time. The embedding model is also locked after the first successful embedding, so a wrong value costs a full re-embed to undo.
Proposed behavior
Either extend the environment table from #4498 to the AI provider registry and the recommendation settings, or let the provisioning document from #4602 carry settings next to plugins. Both already read secrets by reference, which the provider key needs. The fields that matter: provider type, base URL, capabilities, key, embedding model, and whether recommendations are on.
Two things to keep in mind from the current behavior: recommendations.* only applies after a restart, and the embedding model is locked once something is embedded.
Alternatives considered
A separate config file read at startup, next to the provisioning document: another format to document for the same outcome.
Writing server_settings directly: skips validation and the embedding lock.
n/a — this is a feature idea with no code. The endpoints and setting keys named here were checked against silo-server main at 975bbf4; I reviewed the text and the claims myself.
Confirmations
I searched Ideas and the issue tracker, including issues closed as not planned, and this is not already proposed
reacted with thumbs up emoji reacted with thumbs down emoji reacted with laugh emoji reacted with hooray emoji reacted with confused emoji reacted with heart emoji reacted with rocket emoji reacted with eyes emoji
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
Problem
Silo can take server settings from the environment (#4498) and install plugins from a provisioning file (#4602). AI providers and recommendations are in neither, so a deployment kept in git still needs a person in the admin UI, or a script.
Turning on recommendations means four manual steps: connect an embedding provider (type, base URL, key, what it serves), choose its model, enable recommendations, restart. The provider key already comes from a secrets manager in my setup, and the rest is configuration that belongs next to the compose file.
PR #4498 says that variants reading a secret from a file were left out of that change. A provider key is exactly that kind of setting, and the provider registry from #3106 has no way in from outside the admin pages.
Who it affects
Operators who deploy Silo from Compose, Kubernetes or NixOS and keep the rest of the server in version control. The same case as #4594 for transcode node keys.
Current workaround
Script the admin API after first start:
POST /api/v2/admin/ai/providers(with/checkand/{id}/models), thenPUT /api/v2/admin/settings/recommendations.embedding_provider_id,...embedding_modeland...enabled, then restart. This works but needs an admin API key as a deployment secret, can only run once the owner account exists, and reports failure in logs rather than at deploy time. The embedding model is also locked after the first successful embedding, so a wrong value costs a full re-embed to undo.Proposed behavior
Either extend the environment table from #4498 to the AI provider registry and the recommendation settings, or let the provisioning document from #4602 carry settings next to plugins. Both already read secrets by reference, which the provider key needs. The fields that matter: provider type, base URL, capabilities, key, embedding model, and whether recommendations are on.
Two things to keep in mind from the current behavior:
recommendations.*only applies after a restart, and the embedding model is locked once something is embedded.Alternatives considered
server_settingsdirectly: skips validation and the embedding lock.recommendations.enabled.AI harness
pi 1.5.2
AI tool(s)
pi coding agent (file reads, shell, web fetch)
AI model(s)
deepseek-v4-1-flash-1017 via Mistral
AI involvement
AI-assisted
Independent or adversarial review
n/a — this is a feature idea with no code. The endpoints and setting keys named here were checked against silo-server main at
975bbf4; I reviewed the text and the claims myself.Confirmations
All reactions