AVP Compression Specification

Overview

AVP supports optional compression using zstd (Zstandard).

Compression Levels

Level	Zstd Level	Use Case	Trade-off
fast	1	Real-time	Fastest, lowest ratio
balanced	3	Default	Good balance
max	19	Archival/batch	Slowest, best ratio

When to Compress

Compress:

• Embedding payloads > 1KB
• Network-constrained environments
• Batch transmission

Don't compress:

• Hidden state or KV-cache payloads (zstd achieves only 1-7% on random tensor data)
• Payloads < 1KB (overhead exceeds savings)
• Ultra-low latency requirements

Typical Results

Embeddings (compressible -- repeated patterns in floating point):

4096-dim float32 embedding:

• Uncompressed: ~16KB
• Compressed (level 3): ~11KB
• Savings: ~30%

Hidden states and KV-cache (effectively random data):

4096-dim float32 hidden state:

• Uncompressed: ~16KB
• Compressed (level 3): ~15KB
• Savings: 1-7%

For latent communication, the primary value is skipping autoregressive generation, not bandwidth reduction. Compression is most useful for hidden state payloads.