How Core Works¶

How StrataRouter Core routes queries using hybrid semantic search.

What is StrataRouter Core?¶

StrataRouter Core is the semantic routing engine that decides which handler should process a given query. It combines multiple algorithms for maximum accuracy and speed.

Three-Stage Pipeline¶

Every route call passes through three stages:

flowchart TB
    QE["Query Embedding"]

    subgraph ROUTER["Router Engine"]
        direction TB
        HNSW["HNSW Index"]
        HYB["Hybrid Scorer"]
        CAL["Calibrator"]
        HNSW --> HYB --> CAL
    end

    RESULT["Route Result"]

    QE --> ROUTER --> RESULT

Components¶

1. HNSW Index — Fast approximate nearest-neighbor search. Sub-millisecond query time. Hierarchical navigable small-world graphs with O(log N) complexity.

2. BM25 Scorer — Keyword-based sparse matching. Handles exact term matches. TF-IDF with length normalization. Discriminates between similar semantic routes.

3. Pattern Matcher — Deterministic string and regex pattern matching. Fast rule-based routing for known phrases. Override any semantic decision.

4. Score Fusion — Weighted combination of all signals. Learned weights: Dense (64%), Sparse (29%), Rule (7%). Sigmoid normalization.

5. Calibration — Isotonic regression per route. Maps raw fused scores to true probabilities. Expected Calibration Error (ECE) < 0.03.

Performance Summary¶

Metric	Value
P50 Latency	`<2ms`
P95 Latency	`<6ms`
P99 Latency	`<10ms`
Throughput	10,000+ req/s
Memory	~64MB (1K routes)
Accuracy	95.4%

Use Cases¶

StrataRouter Core is the foundation for customer support routing, intent classification, agent selection in multi-agent systems, query categorization, and any workflow requiring fast, accurate decision-making.

Next Steps¶

Routing Engine

Deep dive into the three-stage pipeline.

→

HNSW Index

Vector search internals and tuning.

→

Algorithms

BM25, calibration, SIMD, score fusion.

→

Performance

Benchmarks and optimization guide.

→