← NovaCortex

Retrieval beats dumping your KB into context

NovaCortex fetches only the few facts a question needs. Dumping the whole knowledge base into every prompt burns tokens, money, and (with a hosted provider) ships all your data every time. Same questions, same model — here is the measured difference.

−98.8%
input tokens at a 26K-token KB
~65×
cheaper per query at that size
100%
retrieval recall@5 (with 518 distractors)
flat
retrieval tokens as the KB grows
26K18K9K0~2K KB~10K KB~26K KBknowledge-base size (tokens)dump 25,675 tokretrieval 305 tok (flat)
Dump whole KB into contextNovaCortex retrieval (top-5)

The numbers

KB sizeDump tokRetrieval tokReductionDump costRetrieval cost
42 facts (~2K)2,02830585%$0.0048$0.0009
220 facts (~10K)10,15730597%$0.0231$0.0009
560 facts (~26K)25,67530598.8%$0.0581$0.0009

15 questions, gpt-4o-mini, top-5 retrieval, costs for the full 15-question run. Answer accuracy (LLM-judged against reference answers): dump 100%, retrieval 93%. Retrieval recall@5 = 100% even with 518 unrelated distractor facts in the store — NovaCortex surfaced every fact the questions needed.

How we measured

Two ways of giving the model the project knowledge: dump the whole KB into the prompt every query, or store it in NovaCortex and retrieve only the top-5 relevant memories. We score retrieval separately with a recall@K metric, so the rare answer miss is shown to be an LLM/grader limit (it hits the dump baseline equally), not a retrieval failure. Fully reproducible: node scripts/benchmark/run.mjs.

And it's a privacy win, not just a cost one

Dumping the whole KB ships all of your knowledge to the model provider on every query. Retrieval sends only the few relevant snippets. With NovaCortex self-hosted, the knowledge base never leaves your infrastructure; only the minimal retrieved context goes to the model.

See pricing →View on GitHub