English to SQL plus: 2 layers retrievals, 6 channel embedding. Not a simple Cosine
I got tired of text-to-SQL tools that let the model emit raw SQL against a live DSN.
L1
Dense cosine (plus exact/typo name-link)
Schema names — tables, aliases, long columns
L2
VectorPrism
Schema chunks — 1024d tensor, six channels per chunk (dense core + relational + disentangled + hyperbolic + identity + causal). Query is encode_query → PSM search. Joins still only from declared join_paths.
Prism-Eval is open-source adversarial CI for AI extraction agents. Catch digit drops, prompt injections, and OCR drift before poisoned tool calls reach your engines. G4 false-accept invariant · JUnit/SARIF · Apache-2.0. pip install prism-eval.
