---

11.4 Procedure A: same-dimension switch ¶

Use when: the new model produces vectors of the same dimension as the old one. Examples:

• Stay at 1536; switch from text-embedding-3-small to text-embedding-3-large@1536 (OpenAI supports a dimensions parameter that truncates the larger model’s output).
• Swap to a self-hosted 1536-dim model.
• Switch from one 768-dim model to another 768-dim model after a prior dim-change.

No code or schema changes are required. The whole procedure is config + restart.

1. Back up the database.

   sudo systemctl stop ownsona.service
   pg_dump -h localhost -U ownsona ownsona > /var/backups/ownsona-pre-reembed-$(date +%F).sql
   sudo systemctl start ownsona.service
   

   (Or rely on your nightly backup — but a fresh dump immediately before re-embed is safer.)

2. Edit the deployed application.ini at <tomcat>/webapps/ROOT/WEB-INF/backend/application.ini:

   EMBEDDING_PROVIDER   = openai
   EMBEDDING_MODEL      = text-embedding-3-large
   EMBEDDING_DIMENSIONS = 1536
   EMBEDDING_API_KEY    = sk-...                # if it changed
   EMBEDDING_ENDPOINT   = https://api.openai.com/v1/embeddings
   
   REEMBED_ON_STARTUP   = true
   

3. Restart the service.

   sudo systemctl restart ownsona.service
   

4. Watch the log.

   journalctl -u ownsona.service -f
   

   You’ll see, in order:

   migrator: db_version at N, target N, nothing to apply
   record_migrator: done upgraded=0 failed=0
   reembed: starting active_provider=... active_model=... dims=...
   reembed: progress count=50
   reembed: progress count=100
   ...
   reembed: done count=N
   ApplicationIniWriter: set REEMBED_ON_STARTUP = false in <path>
   

   The server is serving requests during the walker’s run. Recall on rows already updated uses new-model vectors; recall on rows not yet updated uses old-model vectors. Within either cohort, results are correct; across cohorts, scores are not meaningfully comparable. For a single-user store this window is minutes, not hours.

5. Verify and propagate.

   SELECT embedding_provider, embedding_model, count(*)
   FROM memories
   GROUP BY 1, 2;
   

   Should show one row, with the new provider/model and a count matching count(*) overall.

   The walker auto-flips REEMBED_ON_STARTUP=false in the deployed application.ini, but the source tree still has it at true. Before your next WAR build, update src/main/backend/application.ini (and the example file) to match, so the next deploy doesn’t accidentally re-trigger the walker.