Performance and Troubleshooting EN

Performance and troubleshooting

한국어 · Home

Decide whether to enable Laya

Initial M4 Pro/24 GiB measurements, not promises for every machine:

Path	Observation
Repository keywords only, 6-repository fixture	About 11.5 MiB peak process RSS
Repository preview with Laya	About 1.40 GiB RSS; model startup about 942 ms
Mixed 24-request preview set with Laya enabled	Median 39.669 ms, p95 54.040 ms, excluding load; includes skipped inference
Pure-Go shortlist, 1,000 synthetic repositories	About 0.17 ms/request; excludes index construction and inference
Code reranking, 8 candidates in development retrieval tests	Roughly 1.6–1.8 seconds added

On easy synthetic repository fixtures, lexical Top-1 was 18/18 positives; Laya's raw choice was correct for 8/15 judgments and accepted suggestions were zero. Zero coverage does not establish high precision. No actual Codex savings have been demonstrated. Start lexical, then compare on labeled tasks representative of your work.

Common problems

Symptom	What to check
riidolaya: command not found	Run ./riidolaya or ./bin/riidolaya, or add your installation directory to PATH.
Example catalog/config not found	Run from a laya-tools clone; examples are not in binary-only archives.
Missing model/runtime	Run riidolaya doctor, then riidolaya setup; setup verifies downloads.
Download/checksum failure	Check network and proxy configuration; retry setup. Do not bypass checksum or TLS verification.
Search has no results	Use exact identifiers; check root/ignore rules. Reranking cannot recover absent candidates.
Searching a subdirectory fails	Install rg; Git-root and ordinary-directory file discovery differ.
Model route always uses strong	Inspect abstained/reason; default confidence is 0.9, Korean classification is unvalidated. Do not assume reducing the threshold improves outcomes.
Repo preview returns candidate with --laya	Check reason: unsupported script/metadata or inference unavailable can leave lexical candidates.
Repo preview returns abstain	Could be ambiguous/multi-repo, low confidence, none, or truncation. Shorten metadata and inspect the task.
High memory with warm services	Each process may hold a model. Reuse a process when appropriate; stop it when done.
Core ML initialization fails	The tested dynamic export failed; use default CPU. GPU/ANE benefits are unverified.
Agent cannot parse output	Use --json/JSONL, keep stderr separate, and inspect status rather than assuming success means recommendation.

Measure on your machine

riidolaya bench --iterations 30 --threads 4
riidolaya bench --cpu-profile cpu.pprof --heap-profile heap.pprof --ort-profile ort-trace
go tool pprof -top cpu.pprof
go tool pprof -top heap.pprof
/usr/bin/time -l riidolaya bench --iterations 30

The last command is macOS-specific. Go pprof excludes native/GPU memory; do not report Go heap as total RAM. Native calls may lack useful symbols. ORT profiling itself adds overhead. Keep profiles and private catalogs out of public issues; describe version, platform, command, and a redacted reproducible example instead.

Full measurements · Repository evaluation · Issues

Actions performance replay guide: inspect wall time, CPU time, peak RSS, and decisions for fixed scenarios.

Search SoA optimization and difficulty goldens