Set up recommendations
Silo is in active development before 1.0. These guides are updated frequently.
Recommendations suggest titles from your catalog and each profile’s activity. They are optional and need PostgreSQL with pgvector and an embedding model from a provider connected in AI Services. The rows that show them are set up in Home Sections.
Pick the embedding model carefully: Silo locks it after the first successful embedding, and changing it later deletes every embedding.
Configure the model
Section titled “Configure the model”- Open Admin > Integrations > Recommendations.
- Under Embedding Model, choose the Embedding provider, or choose Connect a provider… to add one. A ChatGPT plan doesn’t include embedding models.
- Enter the Model, or pick one of the provider’s suggestions. The field saves when you leave it; wait for the save result.
- Choose Check Connection. It must report the model’s dimensions; a chat-only model won’t work.
- Under General, turn on Enable Recommendations, then restart the server. Silo applies the recommendation settings only after a restart.
To set these from your deployment instead, add three variables to the Silo service’s environment: SILO_RECOMMENDATIONS_ENABLED=true, SILO_RECOMMENDATIONS_PROVIDER set to the id of a provider that serves embeddings, and SILO_RECOMMENDATIONS_MODEL set to the model. Use a provider your deployment document declares: one added in the admin pages gets a generated id that the admin pages don’t show. Enable Recommendations, Embedding provider, and Model then show read-only on this page, and Check Connection still works.
The variables take effect on the start that reads them; no second restart is needed. If the provider isn’t connected or doesn’t serve embeddings, Silo doesn’t start and logs an error such as the recommendation embeddings provider "local-embed" is not connected. Give every API server the same values, because the last one to start wins.
A hosted embedding model receives catalog text from Silo, so review the provider’s charges and data policy first. Leave the advanced ranking settings at their defaults until you have results to judge.
With Enable Recommendations off, no profile gets personal rows. For You still shows rows every profile gets, such as Highly Rated in Your Library and Recently Added. On Home, the Recommended for You shelf shows shared picks under their own row name, or is left out when there are none.
Generate and inspect results
Section titled “Generate and inspect results”In Job status, run Embeddings first. After it finishes, run Taste Profiles, Co-Watch Matrix, and Recommendations in that order, checking each for errors.
Open For You with a profile that has watched, liked, or favorited a few titles. A new profile sees only rows every profile gets, such as Popular on This Server and Highly Rated in Your Library, until Silo learns what it likes; in the web app, Pick what you love gives it a start. Silo then updates a profile’s personal rows in the background after it watches or rates something on any device, without waiting for the scheduled jobs. Every row respects the profile’s library access and rating limit.
If a job fails, check the provider in AI Services, the model, and the job’s error.
The Schedule section sets when jobs run again. Change it after a manual run works.
Change the embedding model
Section titled “Change the embedding model”The Embedding Lock card shows the model Silo is locked to. To move to a different model:
- Choose an Embedding provider that serves the new model, and enter the new Model. If the variables set them, change
SILO_RECOMMENDATIONS_PROVIDERandSILO_RECOMMENDATIONS_MODELinstead. - Restart the server.
- Choose Reset embeddings and confirm. This deletes every embedding and personal recommendation row and clears taste profiles. Personal recommendations and semantic search stay unavailable until the embeddings are rebuilt. If an embedding job is running, Silo refuses the reset until it finishes.
- Run Embeddings again. Taste profiles rebuild as their titles are embedded.
To keep your embeddings, switch only the provider: choose another Embedding provider that serves the same model with the same dimensions, choose Check Connection, and restart once it passes. When SILO_RECOMMENDATIONS_PROVIDER sets the provider, change that variable instead, restart, and then choose Check Connection. If the provider is only down for a while, wait for it to come back rather than resetting.