Installed the library and used it to index helpdesk tickets and replies, retrieve passages, and attach record citations. It refreshes a ticket when the ticket or a reply changes and removes it when that model is deleted. Conversation history, a score gate on ask, and a hard stop when nothing relevant is stored were missing, so those were wrapped in application code. Migrations are stubs that must be published, the default store is in-memory, and the offline embedder never ranked a real question above the acceptance floor.
- What worked
- Eloquent embedding, a database-backed vector store, provenance on search hits, and model hooks for save and delete matched the tests once a scripted model and a very low score floor were used. Citations named the ticket, follow-up turns stayed on one conversation id, and an edited ticket showed up in the next excerpt.
- What got in the way
- Ask has no threshold, so retrieval had to run twice. Empty context still reaches a model unless the caller refuses first. There is no conversation store. Page numbers stay on document sections and are not copied onto chunks. The bundled embedder hashes whole texts, so overlap with a question still scored below zero. Schema files do not load until published. Encryption and the key store needed extra choices even though ticket text is already stored in plaintext. Bulk deletes skip model events, so they leave stale vectors.