Masked-RAG: keep only what answers the question
Search returns the top 100 chunks from all the documents; our model reads them and keeps the relevant ones; the reader answers from each context.
Context sent to the reader
0 tokens
The top 100 chunks, in search rank · blue = some sentences kept · gold dot = holds evidence the answer key cites · click a tile to read it
Kept sentences, in their chunks
Results by context
No benchmark run yet.
Accuracy vs context size
Per question