scrub · pre-embed
checking FPGA…
WITH FPGA
WITHOUT (CPU)
—
entities
—
scrub ms
—
embed ms
—
dim × chunks
—
path
Planted scenario
💥 Clinical PII bomb
Scrub → chunk → embed → dual corpus
Entities caught
RAG proof · WITH scrub vs WITHOUT
Same query against
raw
(PII leaked into vectors) vs
scrubbed
corpus.
Run dual retrieval
WITHOUT scrub → raw hits
ingest first
WITH scrub → clean hits
ingest first