| name | biocheminsight-vision-models |
| description | Use when a BioChemInsight vision model needs to classify candidate structure images, reject Markush or fragment noise, or extract a compound ID from a highlighted structure image. |
| model | inherit |
| effort | high |
| context | auto |
BioChemInsight Vision Models
Use this skill for runtime vision-model tasks.
This skill inherits shared runtime conventions from biocheminsight-model-common and only overrides vision-task-specific behavior.
Scope
- Structure candidate filtering before OCR/SMILES/ID extraction
- Conservative complete-compound gating
- Compound-ID extraction from highlighted structure images
- Visual audit of requested assay cells, including visible units and symbol/grade descriptions
Runtime rules
- Be conservative: cropped or ambiguous candidates must not pass as complete compounds.
- Any variable-group placeholder means Markush, not a complete compound.
- If a bond, atom, ring, label, or substituent is clipped by the image boundary, treat it as incomplete.
- If the image is mainly text, arrows, legends, tables, or multiple unrelated objects, treat it as noise.
- Respect repository strictness policy:
strict, balanced, or permissive.
- For assay-cell audit, do not rewrite tables; only verify requested cells and report visible
value, unit, and symbol/grade description.
- For structure-page detection, emit exactly one
decisions item per provided page.
- For structure ID extraction, use
COMPOUND_ID="None" only with ID_SOURCE="none" and evidence explaining why no reliable ID is visible.
Runtime references
references/classify_structure_candidate_prompt.mdreferences/classify_structure_crop_check_prompt.mdreferences/classify_structure_border_review_prompt.mdreferences/detect_structure_pages_prompt.mdreferences/review_structure_pages_prompt.mdreferences/structure_to_id_prompt.mdreferences/review_structure_primary_id_prompt.mdreferences/runtime.jsonreferences/output_schemas.json
