Draftbox uses up to 90% less compute than comparable platforms. Not by cutting corners — by being smarter. Binary vectors, bloom deduplication, delta processing. The same work, a fraction of the footprint.
Most platforms use HNSW graphs for vector search — an approach that requires the entire index to live in RAM. For 1 million pages, that means ~8 GB of memory constantly consuming energy. Draftbox uses binary Hamming vectors instead: 128 bytes per vector, 16× smaller, fitting entirely in CPU cache. The same search, a fraction of the power.
Others (HNSW in RAM)
~8 GB RAM
Draftbox (binary)
128 MB (CPU cache)
Others (full re-ingest)
100% of blocks
Draftbox (delta + bloom)
~10% of blocks
Others
Full pipeline every time
Draftbox (AuraCache)
~50ms, $0
16× smaller index than HNSW. Fits in CPU L2 cache instead of RAM. Sub-millisecond scan from NVMe. Less memory, less energy, same precision.
30% of content is duplicates. Bloom filters skip them before any LLM call. Zero wasted compute on blocks the system has already seen.
When content changes, only the changed blocks are re-processed. No full re-ingestion. No redundant embedding calls. The minimum work, every time.
If someone already asked a similar question, the cached answer is returned in 50ms. No LLM call. No reranking. No embeddings. The greenest compute is the compute that never happens.
When two team members ask similar questions, the second gets a cache hit. One computation serves many. Less compute, less cost, less energy.
Vercel Edge + Neon serverless + Upstash KV. Compute scales to zero when idle. No always-on servers burning power for nothing. You use it, it runs. You don't, it sleeps.
The infrastructure aligns with ISO 14001 (Environmental Management), ISO 14064 (GHG Emissions), ISO 50001 (Energy Management), and ISO 14040/14044 (Life Cycle Assessment). Not certified. But the design respects the principles.