Step B: Deterministic Pipeline
Same input → same output, every time
Test Coverage
0
Passed
0
Failed
0
Warnings
📋 Step B: Pipeline Determinism (Sage-Validated)
🎯 Philosophy: Extract Once, Process Deterministically
LLM extraction can vary (acceptable). What matters: same extracted data → same downstream processing every time.
3.1 Create Extraction Snapshot
Store conversation → Extract ONCE → Save as immutable snapshot
3.2 Embedding Pipeline Determinism
Process snapshot twice → same embedding count, no duplicates
3.3 Graph Ingestion Determinism
Process snapshot twice → pass2 creates ZERO nodes (deduplication working)
3.4 Recall Stability
Same query on same graph → identical results returned
3.5 Global Deduplication
Verify no duplicate fingerprints exist across entire graph
🎯 Step B: Pipeline Determinism
New Philosophy (Sage-validated): Extract once, process deterministically.
What we're testing:
- Extract intelligence ONCE → create immutable snapshot
- Process snapshot TWICE through embedding pipeline
- Verify: same embedding count, no duplicates created
- Process snapshot TWICE through graph ingestion
- Verify: second pass creates ZERO nodes (dedup working)
- Run same query twice → verify identical recall results
- Global check: no duplicate fingerprints in entire graph
Key insight: LLM extraction can vary (acceptable). What matters: downstream processing is bulletproof deterministic.
Click "Run Step B Test Suite" to validate pipeline determinism.