Benchmarks
Every number on this page came from a real battle-test against a live local Zerocache instance with real provider keys — not a synthetic benchmark.
A support-triage agent suite, re-run
A real ReAct tool-calling agent (search_kb / get_order /resolve_ticket over in-script data) resolves a 3-ticket suite — 10 model calls plus 9 tool calls per run — against the completion cache, live-smoke-tested through Gemini's OpenAI-compatible endpoint with a real key.
The demo/completion-cache/harness (battle-test.mjs and agent.mjs) reproduces this. The largerdemo/agent-showcase/— an architect, three parallel coders, a Claude reviewer over /v1/messages, and a fixer shipping a rate-limiting feature — is built to run cold, then cached, then reworded (semantic tier); run it yourself for a live trace.
Re-indexing an edited corpus
A 9-document corpus, re-indexed after editing one document and adding another. The other 7, byte-identical between versions, were served from cache for free.
Miss vs. hit latency
Measured in the Mastra agentic battle-test, same corpus, before and after caching.
Ingest (per doc) — 196× faster
Query — ~600× faster
Consumer battle-tests — three independent frameworks
LangChain/TS (a raw embedding client, OpenAI) found the bare-string input gap and the first measured coalescing number. LlamaIndex/Python (Gemini, a second language ecosystem) confirmed the fixes held against a second provider. Mastra (an agentic pipeline — an LLM deciding which tool to call, not a direct embedding client) confirmed the same behavior holds when Zerocache is several layers removed from the caller, and exercised theDELETE routes end-to-end for the first time.