Benchmarks & Profiling
Runtime latency, memory overhead, and throughput measurements
Measurements in Preparation
Formal real-device benchmark measurements for AMV-AIChain are currently being compiled. The library is a pure-Python / pure-ESM zero-dependency runtime; its performance is dominated by the upstream LLM backend (llama-server, OpenAI-compatible endpoint) and network round-trip to that endpoint, not by AIChain's own pipe overhead.
Planned measurement axes: LCEL chain end-to-end latency (ms), SQLiteVectorStore cosine search latency (ms, 1K–50K vectors), StateGraph state-transition overhead (µs), and peak RSS delta during chain execution on Android Termux ARM64.
Planned Measurement Conditions
| Metric | Planned Target Device | Conditions | Status |
|---|---|---|---|
| LCEL chain latency (prompt→parse) | Galaxy S25 / A35 (Termux) | Local llama-server, 4 threads, 128-token output | Pending |
| SQLiteVectorStore top-5 search | Galaxy S25 / A35 (Termux) | 10K 384-dim vectors, in-memory :memory: DB | Pending |
| StateGraph node-transition overhead | Galaxy S25 / A35 (Termux) | 5-node graph, 100 iterations, Python 3.12 | Pending |
| Package import RSS delta | Galaxy S25 / A35 (Termux) | Fresh Python process, stdlib-only baseline | Pending |