- compare/route: LLM-only + embedding fetch run in parallel (phase 1);
basicRag + graphRAG run in parallel after retrieval (phase 2).
Reduces compare wall time ~50% vs sequential.
- benchmark/route: all 10 samples run via Promise.allSettled (parallel);
within each sample same 2-phase parallel structure.
Reduces benchmark wall time from N*LLM_time to ~1*LLM_time.
- retrieval.ts: add 256-entry in-process embedding cache keyed by
normalized query text — eliminates repeated HF API round trips.
- llm-providers.ts: cache OpenAI client instances per (baseURL, apiKey)
pair — eliminates re-instantiation and dynamic import overhead on
every callLLM() invocation.
Add .gitignore, dataset metadata, retrieval layer, and latest web/graphrag updates