The Agent Skills Directory

[Data Processing Considerations]: The skill documents functions like AI.AGG, AI.CLASSIFY, and AI.GENERATE which process unstructured data from sources such as Wikipedia comments, user reviews, and emails. When processing untrusted external data, users should consider implementing boundary markers or prompt delimiters as a best practice to manage potential injection risks.
Ingestion points: Data enters through arguments in functions like AI.AGG(input, ...) in references/ai_agg.md, AI.CLASSIFY(input, ...) in references/ai_classify.md, and prompt interpolation in references/ai_generate.md.
Boundary markers: Examples typically use direct concatenation (e.g., 'Summarize: ' || content); adding explicit delimiters is a recommended security enhancement for production use.
Capability inventory: The functions operate within the BigQuery SQL environment, which has permissions to read from and write to datasets and access Google Cloud Storage objects via OBJ.GET_ACCESS_URL.
Sanitization: The documentation focuses on functional usage; users should apply standard data validation and sanitization when handling external inputs.
[Cloud Storage Integration]: The skill provides examples of accessing multimodal data (images and PDFs) stored in Google Cloud Storage using OBJ.GET_ACCESS_URL. The references target official Google sample data repositories and represent standard intended functionality.
Evidence: References to gs://cloud-samples-data/bigquery/tutorials/... are found in references/ai_generate.md and references/ai_similarity.md.

bigquery-ai-ml