Interactive demo · Adversarial benchmarks

VectorPrism

Multi-channel retrieval that recovers evidence dense cosine buries — causal, hyperbolic, and relational signals in one 1024d tensor.

pip install vectorprism GitHub repo Benchmarks markdown Try the demo ↓

Adversarial finance pack

Calibrated hard pack: dense Miss@10 ≈ 93%. Multi-channel Stage-1 expansion + Stage-2 fusion recovers the buried gold documents. Numbers from committed artifacts under demos/finance_demo/results/hard_eval/.

7.1%
Dense R@10
100%
Multi z-score R@10
78.6%
RRF (prod-realistic)
13/13
Recovered dense misses
85%
Auto-graph recovery
~1.8ms
Search @ 1000 docs
ConfigRecovered@10Full R@10Visual
±0.4pp
OOD ΔP@10 (soft packs)
85%
Recovery @ +400 noise edges
100%
After 40% edge dropout
13/13
On 1000-doc scale corpus

Interactive recovery demo

Pick a real adversarial query. Left = dense ranking (gold buried). Right = multi-channel ranking from the measured scorecard (gold recovered into top-10).

Dense rank of gold: — · Multi recovered@10: —

1. DenseTop-K semantic seeds
2. ExpandCausal · taxonomy · predicates
3. ScorePer-channel hop maps
4. FuseZ-score or RRF
Dense top-5
    Multi top-5

      Why dense fails — and channels win

      Cohesive distractors share vocabulary with the symptom. Gold root-cause / sub-clause / predicate docs sit at ranks 12–200 until Stage-1 structure brings them into the pool and Stage-2 topological bonuses overcome the cosine gap.

      ChannelFailure mode it fixes
      Causal DAGSymptom neighborhoods ≠ root cause
      Hyperbolic taxonomyParent/sibling policies outrank deep child clauses
      Relational predicatesFuzzy match beats exact amount / callback / timezone rules

      Reproduce from git

      git clone https://github.com/insightitsGit/VectorPrism.git
cd VectorPrism
pip install -e ".[all]"
vectorprism pilot-check

# Scorecards (Docker recommended on Windows for sentence-transformers)
docker compose run --rm vectorprism python demos/finance_demo/run_multichannel_recovery.py --skip-train
docker compose run --rm vectorprism python demos/finance_demo/run_post_validation.py
docker compose run --rm vectorprism python demos/finance_demo/run_robustness_validation.py

      Artifact index: RESULTS_INDEX.md · Technical report: TECHNICAL_REPORT.md