RAG & Semantic Routing, before ANN.
Route and filter context deterministically in a symbolic semantic domain.
Pre-ANN semantic routing
Lower retrieval noise
Lower IO and compute
Predictable system behavior
Why retrieval pipelines get noisy
Most RAG systems rely on approximate nearest-neighbor search over continuous embeddings. At scale, retrieval becomes noisy: candidate sets are large, routing is probabilistic, and small embedding shifts can change results. This makes context selection harder to control, audit, and stabilize over time.
Symbolic routing as a first-stage gate
SEMQ enables routing and filtering in a symbolic semantic domain before running ANN. Queries and documents are mapped into compact symbolic codes that preserve semantic structure. Routing decisions become cheaper, more stable, and easier to reason about.
Pre-ANN Semantic Routing
Partition and route queries before vector search.
Lower Retrieval Noise
Smaller candidate sets, cleaner context.
Lower IO and Compute
Less bandwidth, fewer scans, cheaper pipelines.
Predictable Behavior
More stable routing under drift and model changes.
How it fits into a RAG pipeline
embed(query) → semq.encode → route(code) → shortlist ann_search(shortlist) → rerank → build_context
Evaluation and experiments
Routing effectiveness, candidate reduction, and downstream answer quality are evaluated under controlled conditions.
View benchmarks“Retrieval systems need routing that remains stable under change.”