Awesome Search KG

Hev

2 min read

Hev

Hev writes at hev mind (https://hevmind.com), a personal site mixing technical writing on search and AI tooling with interactive data essays, and publishes code at https://github.com/hev. The name is a handle rather than a stated full name, and no affiliation is given on the site; the writing is first-person and independent, and it advertises a small line of self-hosted AI tools under the same hev name.

Contribution to This Vault

Two things make the piece worth keeping. First, the result: an untuned general decision model lands within 0.003 mean nDCG@10 of the best purpose-built reranker in the comparison, which is evidence about the category rather than about one vendor. Second, the method — the write-up reports a determinism re-run, a position-bias check by order reversal, a random-permutation floor, and an ablation of the question shape, then states the cases where its own approach loses. That set of checks is a reusable template for evaluating any reranker.

The author is explicit about the limits: public benchmarks mean possible contamination for every model in the table, rerank depth was 30, and latency was measured from a laptop with network included.

Graph View

Backlinks