---

name: retrieval-optimization description: "Operate LimiX sample-retrieval ensemble inference and Optuna search-space tuning." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

Retrieval optimization

Use this sub-skill when you need to explain, preview, or tune the sample-retrieval ensemble path for a local LimiX checkpoint.

What this covers

• sample-retrieval ensemble inference
• Optuna tuning over retrieval parameters
• attention-map generation for retrieval
• safe preview of the retrieval search space

Route away

• Base predictor inputs, tensor shapes, and task-level inference: ../predictor-inference/SKILL.md
• JSON schema, config generation, and config validation: ../configuration-preprocessing/SKILL.md
• Benchmark CLI loops and search_space_sample_num: ../benchmark-cli/SKILL.md

Before you start

• Use a local LimiX checkpoint and a retrieval-enabled config list.
• Retrieval runs are CUDA/GPU-oriented; the shipped project notes that sample-retrieval ensemble inference is intended for very high-memory GPUs.
• Full checkpoint inference is not run here; use the bundled preview script first.

Read first

• references/api-reference.md
• references/workflows.md
• references/troubleshooting.md
• scripts/preview_retrieval_search_space.py

Safety rules

• Do not run Optuna tuning blindly at the default trial count on a large dataset.
• Do not pass a CPU device with use_retrieval: true.
• Do not assume the constructor's attention_score argument overrides the predictor's internal attention flow; the current search loop recomputes through the predictor path.

Quick decision

• Need to inspect or validate ranges without running inference? Use scripts/preview_retrieval_search_space.py.
• Need to tune parameters for a local checkpoint? Follow references/workflows.md.
• Need to understand exact keys or signatures? Follow references/api-reference.md.
• Need a fix for OOM, missing attention, or config mismatch? Follow references/troubleshooting.md.