Budget-aware context compression for RAG & agents. Paste a context, set a token budget, and watch winnow keep the highest-signal chunks. This runs the real Python library in your browser via Pyodide — no server, no API keys.
winnow scores each chunk with BM25 relevance to the query, then selects with Maximal Marginal Relevance under the token budget — keeping relevant chunks while pruning near-duplicates. Kept chunks are highlighted; dropped chunks are dimmed. The Python running here is fetched verbatim from src/winnow.