Free · Local · Not a chatbot — tested on 64,956 social work studies

A tiny free model can out-search the paid AI standard

Meaning-based search finds the study even when the words differ ("kinship care" ↔ "relative caregivers"). We benchmarked the tools that do it. Higher = better search quality (nDCG@10).

EmbeddingGemma + reranker free · 0.3B + 1.5B · runs on a laptop

BEST IN STUDY.846

🏆

Qwen3-Embedding-8B free · 8B · <1 s per search

.842

EmbeddingGemma + light reranker free · fast option · ~2 s per search

.832

OpenAI text-embedding-3-large paid API · flagship

.807

EmbeddingGemma alone free · 0.3B · <1 s per search

.793

OpenAI text-embedding-3-small paid API · the common default

.769

Keyword search what most databases still use

.604

Free & open — nothing leaves your computer Paid commercial API Keyword baseline

🏆 Best search quality

EmbeddingGemma + free reranker (.846) — beats the paid flagship. The reranker re-reads the top results, adding ~14 s per search.

⚡ Fastest free pick

EmbeddingGemma alone (.793) — answers in under a second, still beats the paid default, and runs on a 0.3B model an ordinary laptop can handle.

150 realistic queries · 50,328 blind paired judgments by two independent frontier AI judges · nDCG@10[Perron et al., manuscript in prep]