jpmschweitzer/library-desk
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases 5 Wiki Activity

perf: batch page embeddings via /api/embed and upsert before pruning stale points

Browse Source 
- embed_batch now issues one batched /api/embed request (the old loop
  made one /api/embeddings round-trip per chunk) with a per-text
  fallback that preserves None-for-failed semantics
- update_from_page embeds all chunks in that single call and stores
  them in one Qdrant batch upsert (upsert_points)
- reindex order reversed: upsert new points first, then prune stale ids
  (deterministic uuid5 ids make overwrite safe) so a mid-way failure no
  longer leaves the page with zero vectors
- VectorUpdateSummary gains status (success/partial/failed) and
  chunks_skipped; all-embeddings-failed keeps old vectors and reports
  failure instead of success=True

Measured on a 7-chunk page ingest (local server, llm_tester):
~375ms -> ~181ms median over 3 runs.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01QbFZyDvYksazX6nYQYZ67L
This commit is contained in:
jpmschweitzer
2026-07-14 13:09:37 +02:00
co-authored by Claude Fable 5
parent c35d3c1fa9
commit 8348b4bf92
6 changed files with 246 additions and 38 deletions
Download Patch File Download Diff File Expand all files Collapse all files
CHANGELOG.md
+1
View File
@@ -20,6 +20,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
### Changed (performance)
- **Batched embeddings + delete-last reindex** - `OllamaClient.embed_batch` now sends ONE batched `/api/embed` request (verified against the live Ollama; the old "batch" looped one `/api/embeddings` call per chunk) with a per-text fallback preserving partial-success semantics. `VectorService.update_from_page` embeds all chunks in that single call and upserts them in one Qdrant batch, and the reindex order is reversed: new points are upserted BEFORE stale points are pruned (deterministic uuid5 chunk ids make the overwrite safe), so a mid-way failure can no longer leave a page with zero vectors — the old order deleted everything first. The summary now reports `status` (`success`/`partial`/`failed`) and `chunks_skipped` instead of unconditional `success=True`; a fully failed embedding pass keeps the old vectors and reports failure. Measured on a real 7-chunk page ingest as `llm_tester` against the local server: ~375ms → ~181ms median (3 runs each).
- **Document sync indexing fixed** - `DocumentSyncService._index_vectors` now awaits `ensure_collection` (the coroutine was created but never ran, so fresh tenants had no collection at upsert time), filters out `None` entries from `embed_batch` so one failed chunk embedding no longer aborts the whole document upsert (all-failed still reports failure), and routes the raw `client.delete`/`client.upsert` calls through the async wrapper (`delete_by_filter` / new batch `upsert_points`). Offline unit tests added.
- **Async Qdrant client** - `QdrantClientWrapper` now uses `AsyncQdrantClient` with an explicit timeout (`QDRANT_TIMEOUT`, default 30s). Every vector call previously ran on the synchronous client inside async wrapper methods, blocking the FastAPI event loop for the duration of each Qdrant round-trip. The wrapper API is unchanged (all methods were already `async`), so call sites only gained real awaits. The HybridRAG document leg was moved off the deprecated raw `client.search` onto the wrapper's `search_vectors` (fixing a latent `AttributeError`: it called the nonexistent `ollama.embed_text`, so the leg always reported `failed`), and the health check awaits `get_collections`.