Documentation
¶
Overview ¶
Command regionprobe asks where the gold sits relative to what we returned.
The 848-instance run says HitFile 0.529 and HitRegion 0.375: having found the file, we point at the right code 71% of the time. CoSIL manages 100% and even BM25 manages 82%, so within-file targeting is our worst dimension against every baseline — and it is the one stage no ablation has ever varied.
"Improve it" is not a plan until the miss is decomposed, because the three shapes it can take need opposite fixes:
trimmed — we returned the right symbol and the evidence window cut the
gold lines out of it. A packing problem.
wrong — the right symbol was never in the answer. A selection problem,
and the graph may know where it is: gold is what a solver READ,
and solvers follow call edges.
unindexed — nothing in the index covers those lines. A corpus problem no
ranking can reach.
The split is measured against the SAME response, so the three cannot be confused with one another.
Click to show internal directories.
Click to hide internal directories.