Live demo · 500 FDA drug records

Type any drug query. Watch standard RAG miss the recall.

Five hundred real FDA Drug Enforcement Actions, paired with their approved labels. BM25 retrieves what looks like the question, which is the approval. PortMem retrieves what should govern the answer, which is the recall.

This in-browser slice isolates the authority-tier logic. Production also runs dense embeddings, a supersession resolver, and an LLM-judge re-rank.

query › classify drug → retrieve top-K from 1000 docs → authority-tier rerank → return top-3 with scores

Standard RAG · BM25 over 1000 docs

See top-2 and top-3

PortMem · authority-weighted

See top-2 and top-3

How this demo works

Documents. 500 real FDA drug records. Each pair contains the approved drug label and the FDA Drug Enforcement Action that recalled it.

Live in your browser. When you type a query, BM25 runs over the 500-drug index to identify which drug you mean.

Pre-computed. Top-3 rankings over the 1000-doc corpus (500 labels + 500 recalls). Standard RAG is pure BM25. PortMem is BM25 plus an authority-tier multiplier (×1.85 for recall docs) and a recency boost (×1.10 for post-2020). Both rankings are computed once and cached as static JSON.

Production differs. The deployed PortMem pipeline uses dense embeddings, a supersession resolver, and an LLM-judge re-ranking step. This demo isolates the authority-tier logic so the lift over BM25 is visible.

Verdict. The "DO NOT ADMINISTER" line is regex-extracted from the live FDA recall text in your browser.

Type any drug query. Watch standard RAG miss the recall.

No strong match in the 500-record benchmark.