internal/

Package benchdata loads BEIR-format retrieval benchmark data (the layout CORE-Bench ships: per-repo corpus.jsonl / queries.jsonl / qrels/test.tsv) and manages the on-disk embedding-matrix cache shared by cmd/corebench and cmd/ftdata.

Package enrich generates one-sentence purpose summaries for indexed code symbols using an OpenAI-compatible chat completion endpoint.

Package evalharness drives the shipped binary's `mcp` subcommand over JSON-RPC stdio, exactly like an agent host does.

Package goldensuite is the repeatable regression gate: one definition of "our numbers", run against the shipped binary, comparable across runs.

Package negcases supplies absent-concept queries for false-confidence measurement, and — crucially — verifies per repo that each concept really is absent before it counts as a negative.

Package releasecfg is the single source of truth for the measured release configuration — the exact settings every published gen-eval number was produced with ("tiny cross-encoder + intent ranker + adaptive rerank").

Package selfcases samples label-free retrieval cases from an index: a documented symbol plus the query built from its own docstring.