Reload knowledge_base.json on sibling uwsgi workers after a rebuild

[defangdevs]

Summary

app.py runs uwsgi --lazy-apps --processes 2, so each worker process has its own RAGSystem instance. /trigger-rebuild runs run_rebuild() in whichever worker got the request; that worker's rag_system.rebuild() rewrites the two embedding .npy files and its own self.knowledge_base, but the sibling worker never gets the call.

retrieve() is wrapped in cache_check, which already compared the embedding files' mtimes and called _reload_cache() on a sibling when they changed - but _reload_cache() only reloaded the embedding arrays, never self.knowledge_base, and cache_check never tracked knowledge_base.json's mtime at all. So a sibling worker ends up serving fresh embeddings against a stale knowledge_base list. Since compute_document_scores looks up about/text/path by list index against self.knowledge_base, a rebuild that changes the doc count leaves the sibling returning content for the wrong document, not just outdated content.

Changes

• Track knowledge_base_timestamp alongside the two embedding timestamps (set in __init__ and rebuild_embeddings).
• cache_check now also compares knowledge_base.json's mtime; _reload_cache now also reloads self.knowledge_base from disk, under the existing self._update_lock.
• Two adjacent bugs fixed in the same code path:
  • _reload_cache never took self._update_lock at all.
  • cache_check's error handler called self.rebuild_embeddings() with no argument, though the method requires a knowledge_base arg - this would have raised TypeError if that path were ever exercised.
• get_knowledge_base.py's final write of knowledge_base.json is now atomic (atomicwrites, already a project dependency), since a reader now keys off this file's mtime and could otherwise observe a half-written file mid-rebuild.

No public behavior/API change - this only fixes cross-process cache consistency.

Note on tests

app/test_rag_system.py isn't wired into CI (build.yaml only runs ruff), and has pre-existing, unrelated bugs that predate this change (e.g. setUpClass calls rebuild_embeddings() with no argument, and a self-referential real/fake swap in test_cache_check_rebuild_embeddings_on_error) that make the whole suite error out before it reaches any assertions. I didn't fix those here to keep this PR scoped to the issue; verified this change by manual review plus ruff check/ruff format --check (both green, pinned to the CI version 0.16.5).

Fixes #110

🤖 Generated with Claude Code