8.84M documents on a single 24 GB machine
How a knowledge-heavy team replaced a five-service RAG stack with one Purple8 process — and cut their infrastructure to a single box.
Result
8.84M-record corpus ingested with ~9 GB peak memory — no swap, one machine
The challenge
A knowledge-intensive workload needed retrieval over the full MS MARCO corpus — 8.84 million passages — with graph relationships, vector search, and full-text in one place. The incumbent approach meant a vector database, a graph database, a relational store, an orchestration layer, and a tracing service, each on its own infrastructure.
What we did
The entire corpus was ingested into a single Purple8 process on an Apple M-series machine with 24 GB of RAM. Purple8's bounded memory management kept peak memory usage hardware-limited regardless of corpus size.
The result
- Peak memory: ~9 GB — a stable bounded envelope across the full ingest run that never grew with corpus size.
- RAM returned to the OS after ingest: −3,626 MB; query-time steady state settled around 5.4 GB.
- No swap required on a 24 GB machine.
- One process replaced what would otherwise be five separate services.
The takeaway: the same 24 GB machine can ingest 2M, 5M, or 8M+ nodes with identical peak-memory behaviour. Never project memory by multiplying per-node size by N — that's the old model, not the Cortex model.