project-imaging-x-discovery

Project-imaging-x-discovery

Use this skill for medical imaging dataset discovery and single-dataset download assistance. Default to the local index first. Only widen to the web when the local index is clearly sparse. Resolve scripts/... and references/... paths relative to this skill directory, not the user's working directory.

Files

• datasets.json: local dataset index
• access_rules.json: URL pattern to access/download rule mapping
• scripts/search_datasets.py: deterministic filtering and ranking
• scripts/lookup_dataset.py: deterministic single-dataset lookup for follow-up download help
• scripts/update_index.py: refresh the local index
• references/query-normalization.md: map user language to search arguments
• references/ranking-policy.md: strict-vs-near and ranking rules

Workflow

1. Translate the request into explicit constraints

Extract only the constraints the user actually stated:

• modality
• anatomy / structure
• disease
• task
• dimension
• label requirement
• access requirement
• platform requirement

If you need mapping help, read references/query-normalization.md.

Treat these as hard constraints unless the user phrased them as a preference such as prefer, ideally, or 最好.

2. Run the deterministic search script first

Use scripts/search_datasets.py instead of manually filtering in prose.

Example:

python /path/to/project-imaging-x-discovery/scripts/search_datasets.py \
  --modality mr \
  --structure brain \
  --disease glioma \
  --task seg \
  --dim 3d \
  --label true \
  --access open \
  --prefer-open

The script returns:

• interpreted_query
• strict_match_count
• strict_matches
• near_match_count
• near_matches

3. Present strict matches first

Strict matches satisfy all hard constraints. Near matches are only for fallback.

Never mix them in one main table. When scripts/search_datasets.py already returns ranked strict matches, preserve that order in the main strict table instead of replacing the first rows with hand-picked examples.

If the user asked for open, public, direct download, or 公开可下载:

1. access=open stays in the main strict result set
2. registration and application move to near matches unless the user explicitly accepted them
3. If you still mention a non-open benchmark such as BraTS, state that it is a high-value fallback that does not satisfy the open requirement

When the user did not state an access requirement, rank by the policy in references/ranking-policy.md.

4. Web supplement only when justified

Do one targeted web search only if:

• strict_match_count == 0
• strict matches are too sparse for the user's request
• the user explicitly asks for newer or additional datasets

Skip web search when the local index already answers the query well.

Mark any added items as web supplement. Do not pretend they came from the local index.

If the user asked for a rare modality and the local index has no real coverage for it, say that explicitly and avoid padding the answer with irrelevant near matches from other modalities.

5. Download assistance is one dataset at a time

Do not bulk download.

If the user has already selected a specific dataset, skip broad discovery and switch straight to single-dataset lookup. In that follow-up mode:

1. Use scripts/lookup_dataset.py before any manual prose filtering or broad search
2. strict_matches must contain only the selected dataset when lookup is unambiguous
3. near_matches stays empty unless the lookup itself is ambiguous
4. Do not re-run broad candidate search unless the selected name cannot be resolved
5. Reuse the resolved dataset's exact download_method and auth_instructions; do not replace them with a generic recipe from another platform
6. If the user explicitly says the dataset is already selected, attempt lookup on the exact provided name first even if the token looks generic, short, or resembles a modality/platform label
7. If that lookup is unresolved, stop and ask for the exact dataset name instead of falling back to broad discovery in the same turn

For access=open:

1. Ask for explicit confirmation on one chosen dataset
2. If the user names a dataset, use scripts/lookup_dataset.py to pull its exact metadata first
3. Download only after confirmation
4. Default to the working directory or a user-specified path
5. Warn before downloads larger than 10 GB

For access=registration or application:

1. Do not start the download by default
2. Surface auth_instructions
3. If the user names a dataset, use scripts/lookup_dataset.py before describing access steps
4. Only continue if the user says credentials or approval are already in place
5. Keep the next step focused on login, registration, approval, and the official portal; do not mention unrelated direct-download tools such as wget, curl, or gdown unless the user explicitly asks for alternatives after access is granted

For access=unknown:

1. Say that local index access is unresolved
2. If the user explicitly asks for the exact next step or download path, inspect the official landing page once to resolve it
3. Mark any resolved path as web-confirmed rather than pretending it came from the local index alone
4. If the page still does not make access clear, stop at unresolved

If scripts/lookup_dataset.py returns multiple matches:

1. Do not pretend the selected name resolved uniquely
2. Put the exact family or version candidates in strict_matches
3. Put weaker or derived variants in near_matches
4. Explain the disambiguation in interpreted_query
5. Give the shared next access step only at the family level unless one variant is clearly selected

If scripts/lookup_dataset.py returns resolution_mode=unresolved for a short or generic query:

1. Do not guess the dataset
2. Keep strict_matches empty
3. Explain that the provided name is too short or too generic to resolve safely
4. Ask for the exact dataset name before giving download steps

Output Contract

Use this shape by default:

interpreted_query:
- modality
- structure
- disease
- task
- dim
- label
- access
- platform

strict_matches:
- markdown table

near_matches:
- markdown table
- each item includes why it is near rather than strict

download_next_step:
- what can be downloaded now
- what requires registration or application
- keep it concrete and short; do not add extra option menus unless the user asked for alternatives

Edge Rules

• Do not invent dataset names, access rules, or direct download URLs
• Do not mix strict and near matches into one primary recommendation list
• Do not let a famous benchmark outrank a stricter open result when the user asked for open access
• Do not treat landing pages as direct file URLs without checking
• If the user asks about commercial use or license, say the index does not track that and point them to the official page

Updating the Index

Refresh the local index only when needed:

python /path/to/project-imaging-x-discovery/scripts/update_index.py --github