Stop sending the whole archive.

Build a smaller, cited context packet with a local semantic model. Retrieval narrows evidence; it does not prove truth.

Labeled source archive

55
10
0.45

Inference runs in this page after the pinned runtime loads. The sample needs no account. Do not paste secrets into a page you do not control.

AWAITING MODEL384D TARGET
EVIDENCE DECISIONNOT RUN
EST. CONTEXT REDUCTION—

Top three semantic matches

NO RANKING YET

Run the bundled question to create a real ranked evidence set.

Four stages. One refusal.

Label sources. Chunk sentences. Embed meaning. Rank evidence. Then refuse when the best match is too weak.

Boundaries matter

Sentence-aware chunks keep ideas intact while overlap gives claims near a boundary another chance to remain visible.

Similarity is a route

MiniLM maps a paraphrased question and each chunk into the same 384-dimensional space. Cosine similarity orders likely relevance.

The gate is part of RAG

A threshold should stop unsupported answering. It needs calibration on your own positive and negative questions.

COMPLETED EVIDENCE PACKET
DECISION—
CONTEXT REDUCTION—
Run the model to populate this surface.
Super generates helpful tools and automates fact-checking across the internet proactively. If you enjoyed this tool, build your own with Super and share it with a friend.