DEFT Mining and Embedding Skill

You are the operator of the DEFT embed-then-mine workflow for VCN AOI. Your job is to take a parquet of weak target images (the gap-analysis or routing output) and a source pool, then produce a deduplicated parquet of mined source images that look similar to the targets — ready to feed into the next training round.

The workflow is fixed and deterministic: embed the targets, embed the source pool, then mine nearest neighbours. Each step's output parquet is the next step's input. There is no iterative search, no clustering pass, no human-in-the-loop selection — depth comes from picking the right encoder and the right topn, not from a multi-phase investigation.

The whole skill is a thin wrapper around three direct docker run invocations against the taotoolkit.dataservices image declared in versions.yaml (resolved at runtime — see Setup). The container's entrypoint takes <category> <action> -e <spec.yaml> [hydra overrides...] — pass embedding imageembeddings -e <embeddingspec.yaml> … for embedding and tmm nearestneighbors -e <miningspec.yaml> … for mining. The -e flag points at a YAML that supplies default values for the subtask's schema; anything afterward is a bare Hydra override (key=value) that selectively overrides spec fields per run. (There is no dataset keyword inside the container — that's the TAO launcher's pillar prefix and is dropped here.) Pull the image once if it isn't cached: docker pull "$DSIMAGE" (after resolving $DSIMAGE per Setup).

Schema keys can rename between data-services releases (the RCA skill saw inferencecsv → inferenceresultsdir, outputdir → resultsdir). When in doubt, introspect the actual schema once per image: docker run --rm "$DSIMAGE" embedding imageembeddings --cfg=job and ... tmm nearestneighbors --cfg=job.

Inputs

1. Target parquet — the gap-analysis output, typically mining_gaps.parquet from tao-route-visual-changenet-samples (or gaps.parquet from tao-analyze-gaps-visual-changenet if routing was skipped). Required column: filepath. If label is also present, label-aware filtering during mining is available; otherwise the mining task silently no-ops the filter.
2. Source pool — a parquet of candidate images to mine against, with a filepath column. If the user only has a CSV, convert it to a parquet with the same columns before Step 2. For label-aware filtering, the pool must also carry a label column.
3. Embedding spec file — a YAML containing model, modelpath, batchsize, and (only when modelpath is a TAO .pth/.ckpt) modelconfigpath. Reused across Steps 1 and 2; inputparquet/output_parquet are supplied per run as Hydra overrides. The same spec MUST drive both embedding steps — embeddings from different encoders are not comparable, and mismatched encoders are the most common cause of "the mined images look unrelated" reports.
4. Mining spec file — a YAML containing topn, knnmetric, filterbylabel, and (rarely changed) sourceembedcolumnname/targetembedcolumnname. sourceparquet/targetparquet/outputparquet are Hydra overrides at run time. SigLIP and CLIP embeddings should use knnmetric: cosine. When filterby_label: true but either embedding parquet lacks a label column, the container logs a warning and proceeds without filtering.

Setup

Resolve the concrete taotoolkit.dataservices URI from versions.yaml once at the top of the run, then confirm Docker, the NVIDIA container toolkit, and a GPU are present before doing anything else. A GPU is required for both the encoder forward pass and the cuML/cuDF k-NN search; both steps fail without CUDA.

# Resolve tao_toolkit.data_services → concrete nvcr.io/... URI from versions.yaml
DS_IMAGE=$(python3 -c "import yaml,os; print(yaml.safe_load(open(os.environ['TAO_SKILL_BANK_PATH']+'/versions.yaml'))['images']['tao_toolkit']['data_services'])")
echo "DS_IMAGE=$DS_IMAGE"

docker info > /dev/null && echo "OK: docker"
nvidia-smi > /dev/null && echo "OK: GPU"
docker image inspect "$DS_IMAGE" > /dev/null \
  || docker pull "$DS_IMAGE"

Every host path the container reads or writes must be bind-mounted. The most predictable approach mounts the workspace root with identical paths inside and outside the container, then reuses one $DOCKER alias for the three invocations:

WORKSPACE=<absolute path that contains all parquets, outputs, and the source-pool images>
DOCKER="docker run --gpus all --rm --ipc=host -v $WORKSPACE:$WORKSPACE -w $WORKSPACE $DS_IMAGE"

Do not pass --user $(id -u):$(id -g) — it triggers a getpwuid() KeyError during the transformers import before any work starts. The container runs as root; chown outputs back to the host UID afterward.

Author the two spec files once per iteration, placing them under $WORKSPACE so the -e argument resolves on both sides of the mount; per-run values stay out of the spec and are passed as Hydra overrides. If the source pool is a CSV, convert it to parquet up front (preserving filepath, and label if present). The default embeddingspec.yaml uses model: SigLIP, modelpath: google/siglip-base-patch16-224, batchsize: 64; the default miningspec.yaml uses topn: 5, knnmetric: cosine, filterby_label: "false" (quoted — the schema reads it as a string).

See references/setup.md for the full environment notes, TAOSKILLBANK_PATH handling, the path-mounting rationale, the getpwuid chown workaround, the CSV-to-parquet snippet, and the verbatim spec-file authoring blocks.

Method

Three commands, in order. Each command's output parquet is the next command's input. Run them as plain Bash; the $DOCKER alias from Setup handles the container, GPU, and mounts. Every invocation follows the same shape: -e <spec> for the baked-in defaults, then a handful of Hydra overrides for the run-specific paths.

Step 1 — Embed the target images

$DOCKER embedding image_embeddings \
    -e <embedding_spec.yaml> \
    input_parquet=<target_parquet> \
    output_parquet=<target_embeddings_parquet>

Reads the gap-analysis / routing output and writes a parquet with filepath, embedding, and any extra metadata columns (e.g. label, siamesescore, weakness) carried forward verbatim from the input. Print the output schema (pd.readparquet(...).columns) to stdout so the script-check hook can confirm the embedding column exists.

If you need to override model / modelpath / batchsize for one run without editing the spec, append them as Hydra overrides (e.g. model_path=...).

Step 2 — Embed the source pool

$DOCKER embedding image_embeddings \
    -e <embedding_spec.yaml> \
    input_parquet=<source_pool_parquet> \
    output_parquet=<source_embeddings_parquet>

Same command shape as Step 1, applied to the source pool. Use the identical embeddingspec.yaml as Step 1, and do not override model / modelpath / batch_size differently here — mismatched encoder configs across the two steps produce non-comparable embeddings.

Step 3 — Mine nearest neighbours

$DOCKER tmm nearest_neighbors \
    -e <mining_spec.yaml> \
    source_parquet=<source_embeddings_parquet> \
    target_parquet=<target_embeddings_parquet> \
    output_parquet=<mined_parquet>

For each target embedding, finds the topn closest source embeddings under the chosen metric, deduplicates across targets, and writes a single-column (filepath) parquet of unique mined source paths. The container also drops a miningsummary.txt next to the output parquet with: query count, neighbour count, duplicates removed, and (when label filtering is on) kept-vs-dropped pair counts. Tweak topn, knnmetric, or filterbylabel via inline Hydra override when sweeping (e.g. topn=10) — no need to rewrite the spec.

When filterbylabel=true but one of the embedding parquets is missing the label column, the container logs a warning and proceeds without filtering. If the mined output looks larger than expected or contains cross-label pairs, scan the docker log for that warning before assuming the task did the right thing.

See references/reference-invocation.md for the minimal paste-and-edit end-to-end recipe (resolves $DS_IMAGE, writes both specs, runs all three steps, chowns outputs, and prints row counts) to run as a single streamed Bash block.

Outputs and report

Write everything into a timestamped folder under the experiment / iteration directory. Get the real timestamp by running date +%Y-%m-%d%H%M%S in Bash — do NOT hardcode or guess. If the user specifies a custom output path, use it directly but maintain the same internal layout. The packaging hook adds miningconfig/ and claudesession.jsonl automatically when MiningReport.md is written.

The mined parquet is the artifact downstream training consumes. The two embedding parquets are intermediate but worth retaining — reusable across multiple mining runs against the same source pool, and the only place to look when a "looks unrelated" report needs encoder-level debugging.

See references/outputs-and-reporting.md for the full output-directory layout and the verbatim Mining_Report.md template (Verdict, Inputs, Encoder Consistency, Mining Run, Per-Label Breakdown, Output Sanity, Recommended Actions; keep it 600–1200 words).

Common pitfalls

The most frequent failure is mismatched encoders between the two embedding steps — the single most common cause of garbage mining output; both steps must consume the same embeddingspec.yaml. Other recurring traps: passing --user (the getpwuid KeyError), skipping an embedding step, a missing label column silently no-oping filterbylabel=true, spec files outside $WORKSPACE, unresolved ??? sentinels, TAO checkpoints without modelconfigpath, CSV source pools fed in directly, host/container path mismatches, no GPU, an unpulled or :latest image tag, and topn × Ntargets ≫ source size (expected — report the actual mined count).

See references/troubleshooting.md for the full pitfall list with the exact errors, causes, and fixes.

Execution Order

1. Resolve DSIMAGE from versions.yaml (images.taotoolkit.dataservices), then run docker info, nvidia-smi, and docker image inspect "$DSIMAGE" (pulling if missing) once to confirm the environment. Abort with a clear message if any fail.
2. Run date +%Y-%m-%d%H%M%S to get the timestamp; create <outputdir>/mining_results/<timestamp>/.
3. Write embeddingspec.yaml and miningspec.yaml into the timestamped dir, filling in the encoder choice and mining knobs. Keep these under $WORKSPACE so the -e path resolves inside the container.
4. If the source pool is a CSV, convert to parquet first (preserve filepath and label).
5. Run Step 1 (embed targets) via docker run … embedding imageembeddings -e embeddingspec.yaml inputparquet=… outputparquet=…. Print the output parquet's row count and columns to stdout.
6. Run Step 2 (embed source pool) with the identical embedding_spec.yaml as Step 1. Print output row count and columns.
7. Run Step 3 (mine nearest neighbours) via docker run … tmm nearestneighbors -e miningspec.yaml sourceparquet=… targetparquet=… outputparquet=…. Confirm miningsummary.txt was written next to mined.parquet.
8. Compute the per-label breakdown (Section 5) by joining the target embeddings parquet with the mined output on filepath, if both carry label.
9. Write Mining_Report.md last — writing it triggers the packaging hook, which copies session logs and skill config alongside.