What does this skill do, and when should you use it?
This skill runs the DEFT embed-then-mine workflow for VCN AOI iterations. It reads a target parquet and a source-pool parquet, embeds both with the same encoder, and retrieves nearest source images for each target. The workflow consists of three Docker invocations and produces a deduplicated mined-path parquet plus a mining summary. Docker, nvidia-container-toolkit, and a CUDA GPU are required.
Resolves the tao_toolkit.data_services image from versions.yaml and checks Docker, GPU availability, and the cached image; runs image_embeddings on the target parquet and source pool using one shared embedding_spec.yaml; runs tmm nearest_neighbors with mining_spec.yaml; applies topn, knn_metric, and optional label-aware filtering; writes two intermediate embedding parquets, a mined parquet containing unique filepath values, and mining_summary.txt.
- A VCN AOI engineer needs to expand a real-image augmentation queue from gap-analysis or routing output.
- A data scientist has target weak images and a candidate source pool and needs visually similar source images.
- A vision team wants to sweep topn or label-filtering settings across repeated mining runs.
- An engineering team needs to retain embeddings for diagnosing unrelated mining results or comparing encoder configurations.
How do you install this skill?
- The container runs as root and mounts the entire workspace; manually confirm mount scope, data sensitivity, and overwrite risks before execution.
- The workflow pulls and runs an external tao_toolkit.data_services image; lock and review the concrete image source and credential requirements in versions.yaml.
- If label is missing, filter_by_label silently degrades to unfiltered mining; inspect logs and mining_summary.txt rather than relying on process success.
- Core execution requires a CUDA GPU, Docker, and the NVIDIA container toolkit, with no mainland-China reachability or offline-image plan documented.
- Automatic packaging of claude_session.jsonl and .claude configuration may expose session or environment data; review and sanitize before sharing results.
- Shell / CLI
- Network access
- Local filesystem
Dockernvidia-container-toolkitCUDA GPUPython 3tao_toolkit.data_services container image
Install the selected skill with the repository's documented skills CLI command:
npx skills add nvidia/skills --skill tao-mine-aoi-images --yes
The README does not document a standalone package or a specific manual-copy destination for this skill.
tmp="$(mktemp -d)"
git clone --depth 1 https://github.com/NVIDIA/skills.git "$tmp"
mkdir -p ~/.claude/skills
cp -R "$tmp/skills/tao-mine-aoi-images" ~/.claude/skills/
rm -rf "$tmp"Generated from the source repository and skill path; it copies only this skill's folder. If the author's install steps above differ, follow those first. To scope it to one project, replace ~/.claude/skills with that project's .claude/skills.
How do you use this skill?
Once installed, send your agent any of these to trigger it:
- Run the tao-mine-aoi-images DEFT embedding and nearest-neighbour workflow on my VCN AOI target and source-pool parquets.
Prepare a target parquet and source-pool parquet containing filepath, then create embedding_spec.yaml and mining_spec.yaml under the mounted workspace. Ask a loaded agent something like: “Run the tao-mine-aoi-images DEFT embedding and nearest-neighbour workflow on my VCN AOI target and source-pool parquets.” Before execution, resolve DS_IMAGE and check Docker and nvidia-smi. Run target embedding, source-pool embedding, and nearest-neighbour mining in that order. Both embedding runs must use identical encoder settings; SigLIP or CLIP embeddings should use the cosine metric.
What are this skill's strengths and limitations?
- Provides a fixed, deterministic three-step workflow.
- Supports runtime overrides for topn, distance metric, and label filtering.
- Retains intermediate embeddings for reuse and encoder-level debugging.
- Writes summary metrics including query count, neighbour count, and duplicates removed.
- Requires a CUDA GPU; CPU-only execution is not supported by the documented workflow.
- Requires Docker, nvidia-container-toolkit, and the tao_toolkit.data_services image.
- Different encoders across the two embedding steps make the embeddings non-comparable.
- The supplied material does not document an independent test suite or cross-platform validation.
How does this skill compare with similar options?
Side by side with related skills; every score comes from the same FSRS standard.
| Skill | FS score | Stars | Last updated | License |
|---|---|---|---|---|
| TAO AOI Image Mining this page ✓ NVIDIA · Official | 45 · Not recommended | ★ 3.8k | today | Apache-2.0 |
| TAO VCN Sample Router ✓ NVIDIA · Official | 47 · Use with care | ★ 3.8k | today | Apache-2.0 |
| TAO ChangeNet Root Cause Analyzer ✓ NVIDIA · Official | 39 · Not recommended | ★ 3.8k | today | Apache-2.0 |
| TAO AutoML + DEFT Three-Phase Pipeline ✓ NVIDIA · Official | 51 · Use with care | ★ 3.8k | today | Apache-2.0 |
| TAO DEFT AOI Improvement Loop ✓ NVIDIA · Official | 50 · Use with care | ★ 3.8k | today | Apache-2.0 |
How did FollowSkills review this skill?
The documentation makes the container workflow, workspace data flow, GPU requirement, and Apache-2.0 licensing explicit, with NVIDIA provenance stated; however, Bash execution, root containers, whole-workspace mounting, automatic image pulls, and session/config copying are permitted without user confirmation, least-mount guidance, sensitive-data handling, or robust rollback, so points are deducted.
The three-step order, input/output contracts, pitfalls, and several validation hooks are reasonably clear; however, execution depends on Docker, NVIDIA Container Toolkit, CUDA, an external image, and versions.yaml plus reference files not supplied here. There are no committed key-path tests, and several failures still shift diagnosis to the user, so the static score is capped and reduced.
The audience, AOI/VCN scenario, parquet requirements, filepath/label fields, outputs, and non-iterative boundary are fairly clear; however, the skill is narrow, requires Docker and a CUDA GPU, gives no explicit Chinese-language or mainland-China reachability guidance, and depends on an external NVIDIA registry, reducing environmental fit.
The document is layered into Inputs, Setup, Method, Outputs, Common pitfalls, and Execution Order, and declares author, version, and license metadata; however, version 0.1.0 has no changelog, explicit maintenance owner, or update path, while referenced setup, troubleshooting, and invocation documents are absent from the supplied material and examples/FAQ coverage is incomplete.
The workflow defines a directly consumable mined.parquet and retains intermediate embeddings and a summary for downstream use and debugging; however, output quality depends on encoder choice, data quality, the external container, and GPU availability, with no verifiable representative output. The single-task benchmark is thin and reports only 20% Codex correctness, so effectiveness remains limited under static review.
The files provide concrete commands, output layouts, validation hooks, an evaluation task, and a single benchmark report, offering limited auditability; however, there are no execution logs, CI test suite results, or independently reproducible artifacts, and benchmark coverage is minimal, so key claims remain only partially verifiable.
Open a dimension to read why it scored that way
Evidence confidence：Low — Mostly static review, author material or a limited demo; useful for discovery, not high-risk decisions.
See the full review method →