Block 19 — Production-Observation-Baseline (Tag 0)
Datum: 2026-05-03
Cutover-Commit: 973c670
Branch: feature/block-19-bge-m3-embedder
Diese Baseline dokumentiert den Production-Zustand direkt nach dem
Block-19-Cutover (BGE-M3 Hybrid-Retrieval als Default). Sie dient als
Referenz für Tag-3 (2026-05-06) und Tag-7 (2026-05-10) Observation-
Reviews.
Stack-Konfiguration
| Setting |
Wert |
| use_hybrid_retrieval |
True |
| qdrant_collection |
avs_handbuecher_bge_m3 |
| embedding_model_dense |
BAAI/bge-m3 |
| embedding_model_sparse |
BAAI/bge-m3 |
| rrf_dense_weight |
0.7 |
| rrf_sparse_weight |
0.3 |
| reranker |
mmarco-mMiniLMv2-L12-H384-v1 |
| no_match_threshold |
0.10 |
Re-Verify-Latenzen (Production Hot-Path, drei Test-Set-Queries)
| Query |
Difficulty |
Latenz |
Top-1-Score |
Sources |
Antwort-Länge |
| q003 |
easy |
2.7 s |
0.9548 |
5 |
1171 chars |
| q010 |
medium |
5.3 s |
0.9591 |
5 |
1992 chars |
| q022 |
table |
6.0 s |
0.9939 |
3 |
2383 chars |
Beobachtung: Latenz korreliert mit Output-Länge, nicht mit Retrieval-
Komplexität. Plausibel da 97% Total-Latenz = LLM-Generation
(Phase-3-Eval-Lesson). Pipeline-interne Stages (Embedding 20ms,
Retrieval 8ms, Ranking 37ms) sind subskalig zur LLM-Stage.
Phase-3-Offline-Eval-Referenzen
| Metric |
Block 18 (Baseline) |
Block 19 (Phase 3) |
| R@5 |
0.92 |
0.96 (+0.04) |
| R@10 |
0.92 |
0.96 (+0.04) |
| MRR |
0.7667 |
0.74 (-0.03) |
| avg_retrieval_latency |
155 ms |
140 ms (-15 ms) |
| Tabellen-Queries R@5 |
1.00 |
1.00 |
| Failed |
0 |
0 |
| Regressed (Hit@5↓) |
— |
0 |
MRR-Drift: Isoliert auf q006 (Cross-Encoder-Reranker-Bias bei
duplizierten Inhalten, Phase-3.5-Diagnose). Hit@5 bleibt True,
User-facing-Impact = 0. Mitigation in Block 19.5
(bge-reranker-v2-m3-Upgrade).
Collection-Status (Tag 0)
avs_handbuecher_bge_m3: 242 Points, dense (text-dense) + sparse
(text-sparse) Schema, Status green
avs_handbuecher (Legacy): bleibt 7+ Tage als Recovery-Point
- Snapshot:
backups/qdrant/avs_handbuecher_pre_block19_cutover_20260503-1150.snapshot
(4.3 MB)
Recovery-Plan
| Komponente |
Recovery-Pfad |
| .env |
cp .env.block19_backup .env && make redeploy |
| Collection |
Qdrant Snapshot-Restore aus backups/qdrant/ |
| config.py-Defaults |
git revert <Phase-4-Commit-SHA> (973c670) |
| Code-Pipeline |
git checkout platform/v1.0.0 (Pre-Block-19) |
Observation-Plan
| Tag |
Datum |
Aktion |
| 0 |
2026-05-03 |
Baseline-Capture (diese Datei) |
| 3 |
2026-05-06 |
Mid-Window-Review (p50/p95-Latenz, Error-Rate, no_match-Rate gegen Baseline) |
| 7 |
2026-05-10 |
End-Window-Review + Stichproben-Eval, Cleanup-Entscheidung |