How the Models Were Chosen

Retrieval-quality engineering behind this demo · evidence, and its limits

Benchmarks ← Back to search

Two models decide what this demo returns. An embedding model turns code and questions into vectors, and a cross-encoder reranker reorders the candidates that come back. Both were chosen by measurement on this corpus rather than by benchmark reputation, and both write-ups report the experiments that failed alongside the ones that worked.

Choosing the embedder

Qwen3-Embedding-0.6B · 1024-dim

Three eligible models measured with the corpus, chunking, fusion and reranker held identical. Includes the two constraints that disqualify most candidates before accuracy is considered, and a result that went against the shipping model.

Read the embedder evidence →

Choosing the reranker

Qwen3-Reranker-0.6B-seq-cls

Six rerankers evaluated, a vendor benchmark thrown out after independent verification, and the prompt-format finding worth three orders of magnitude. Also covers why the larger model was rejected despite scoring higher.

Read the reranker evidence →

Why these are separate pages

The two choices are not independent, but they fail differently. A weak embedder can be carried by the keyword legs of the hybrid search, and on this corpus it measurably is. A weak reranker cannot be carried by anything, because it is the last thing to touch the results. Keeping the evidence apart makes it clear which component each measurement is actually about.

← Back to search