BEAVER: Hierarchical Prompt Compression

BEAVER is a training-free, structure-aware prompt compression framework that keeps discourse integrity while delivering extreme efficiency (~26x speedup) on long-context LLMs. Upload your long text or PDF below, adjust compression parameters, and see the results instantly.

Context (long text)

Query / Question

The query guides which pages are most relevant. BEAVER scores each page against your query using semantic + lexical matching.

Tip: Click Auto Suggest to get recommended parameters based on your document length. Shorter texts need fewer pages; longer texts benefit from higher flash_top_k.

Page Size

Tokens per page. Smaller = finer granularity but slower. Default: 64

16 256

Anchor Pages

Force-keep first N pages (title, intro). Preserves structural info. Default: 4

0 16

Flow Window

Sliding window around selected pages for local coherence. Default: 4

0 16

Flash Top-K

Number of highest-scoring pages to keep via semantic+lexical matching. Default: 22

1 64

Semantic Weight

Weight for embedding cosine similarity. Higher = more meaning-based. Default: 0.7

0 1

Lexical Weight

Weight for token overlap. Higher = more keyword-based. Default: 0.3

0 1

Results

Compressed Text