Benchmark Results
Fewer API calls. Lower costs. Same answers.
PRL3 is a drop-in proxy that sits between your application and your LLM provider. It intercepts incoming queries and resolves as many as possible without forwarding them to the LLM. Only queries that genuinely require the model are sent through.
The system learns continuously from live traffic. The more queries it processes, the higher the resolution rate becomes — without any manual configuration.
94% accuracy. 97% → 6% LLM dependency.
A 100-question dataset covering factual, definitional, computational, and multilingual queries was used to measure both answer quality and LLM call reduction across two evaluation passes.
In the first pass (R2), the system learned from scratch — 97 out of 100 queries required the LLM. In the second pass (R3), run after a 70% cache purge, only 6 queries required the LLM. Accuracy held at 94% across both passes.
Independently reproducible. Full install script public.
| Parameter | Detail |
|---|---|
| Accuracy dataset | 100 manually curated questions — fact, definition, calculation, multilingual |
| Throughput dataset | 1,000 queries simulating real chatbot traffic with natural repetition |
| LLM backend | Groq API (cloud) · Local Qwen 2.5 1.5B (air-gapped mode) |
| Semantic verification | MiniLM 384d embeddings · similarity threshold 0.55 for cache admission |
| R2 → R3 protocol | Full learning pass → 70% cache deletion → validation pass on same dataset |
| Hardware | Intel i5-6600 · Quadro K2000 · Ubuntu 22.04 · kernel 5.15 |
| Reproducibility | Full install script at icomnewtechnologies.com/prl3/ |