Skip to content

Name-Only Path Routing: Finding Files Without an Index

Name-only path routing picks directories by name alone, and finds files a fixed lexical query ranks nowhere.

Name-only path routing hands a model the issue text plus a menu of child directory and file names, then walks into the children whose returned probability clears a threshold. It builds no index and reads no file contents. Across 82 audited issues from 11 repositories at pinned pre-fix commits, the walk recovered 0.465 of the files a fix touched within eight candidates. An in-memory SQLite FTS5 ranking reached 0.352 and a fixed full-issue rg replay 0.245 (Bajaj, 2026). The paired gain over FTS5 was +0.113, with a 95% repository-cluster bootstrap interval of 0.053 to 0.168.

Conditions this holds under

The result is narrower than "routing beats grep".

  • The comparison is one fixed query. Each lexical arm gets one non-adaptive shot, and the study "tests an issue's first fixed query, not an adaptive agent" (Bajaj, 2026). On a different task, repository-level code completion, a baseline where the model writes its own ripgrep commands "achieves performance comparable to sophisticated graph-based baselines" (Better Call Grep, 2026).
  • The fix lands in one directory. The subgroup gain over FTS5 was +0.229 across 35 single-directory gold sets and +0.027 across 47 multi-directory sets (Bajaj, 2026). Holding several branches open is what routing does worst.
  • Your own repository is one where it works. Routing beat FTS5 in nine of the 11, tied in one, and lost in one (Bajaj, 2026).
  • File discovery is the thing measured. No issue was resolved in this study, and complete coverage of the annotated lines stayed rare after packing: "14/82 issues for tree, 8/82 for FTS5, 5/82 for rg, and 15/82 for fusion" (Bajaj, 2026).

Why it works

The issue and the source code disagree about words. A bug report describes behavior; the code names identifiers. A full-text query built from that report "may match documentation, changelogs, tests, and unrelated implementation", while the directory structure "supplies an alternate route to candidate files" (Bajaj, 2026). Scoring path names scores a different signal from BM25 over file contents, so the two arms miss different files: routing found 61 gold-file occurrences absent from FTS5's top eight across 40 issues, and FTS5 found 34 absent from routing's.

What the hierarchy contributes is unproven. A flat control showed the same model every eligible path in bounded menus, with no directory hierarchy at all, and scored higher: "An exploratory flat path control reached 0.572 recall while using 24.6 model calls per issue, compared with 8.9 for routing" (Bajaj, 2026). The author says so directly: "The flat result rules out a claim that the hierarchy itself is shown to cause the original gain over lexical search." The flat control is not a clean ablation either, since it "used a different instruction and more calls than tree", roughly 2.8 times as many. What the two arms give is an operating point, not a causal split: flat scored +0.107 above routing at eight files, interval 0.017 to 0.224, for 24.6 model calls against 8.9.

When this backfires

  • Repositories where the walk simply does not work. VS Code scored 0.13 for both arms and Serverless lost to FTS5 at 0.16 against 0.23 (Bajaj, 2026). The paper does not say why; treat the gain as a property of the repository you measure.
  • Latency-sensitive turns. Routing averaged 8.93 seconds and 8.89 model calls per issue, where an FTS5 query averaged 7.5 ms after an 894 ms index build (Bajaj, 2026). The paper's own verdict: "Its 8.93-second mean latency and model charge make a blanket replacement for FTS5 unattractive."
  • Fusing the two candidate lists. Reciprocal-rank fusion raised mean recall to 0.491 but recovered fewer complete gold-file sets than routing alone, 23 against 25. "Rank fusion can displace a tree-selected gold file at a fixed limit; combining candidate lists is not free" (Bajaj, 2026). Keep both shortlists rather than merging them.
  • Branches a name policy prunes. Five of 87 selected cases were dropped because gold paths sat outside the searchable-file policy, "including real source under directories named build" (Bajaj, 2026). A name-driven walk declines exactly those directories.
  • Any claim about fixes shipped. Recall is not repair, and a framework that seeds candidates deterministically before letting an agent revise them does report the downstream number, raising the SWE-bench Lite resolution rate "from 44.00% to 52.33%" over the strongest baseline, with a Qwen3.5 35B repair model (SemNav, 2026).

Example

The measured loop is reproducible without the authors' tooling (Bajaj, 2026). Each menu adds a none_of_these option, and the instruction given was:

Choose the best route toward answer evidence. Directory options summarize descendants, so select a directory when a descendant could answer. Choose none only when the question is unrelated to every option.

The traversal parameters:

Parameter Value
Issue text shown per menu First 500 characters
Descendant names per directory option Up to 12, from at most 3 levels below
Child retention threshold The larger of 0.01 and 0.01 times the best child
Model-call cap 32 per issue
Time cap 180 seconds per issue
Output At most 16 ranked file paths

Path rank is traversal order rather than a multiplied probability, and no issue hit either cap.

Key Takeaways

  • Routing buys +0.113 file recall over a fixed FTS5 query, interval 0.053 to 0.168, for 8.89 model calls and 8.93 seconds per issue. Decide against those two numbers, not the headline.
  • Do not credit the hierarchy. A flat pass over every path scored higher, so the tree is an observed way to spend fewer calls and not a demonstrated cause of the gain.
  • Run the two shortlists side by side. Rank fusion raised the mean and lost complete gold-file sets.
  • Measure on your own repository first. The same method won nine repositories, tied one, and lost one.
  • Nothing here says a fix lands sooner. Complete line coverage reached 14 of 82 issues at best, and issue resolution was never scored.
Emerging — single recent source Last reviewed: 2026-10-02