The Agent Skills Directory

[SAFE]: The skill relies on official and well-known libraries such as pymilvus, transformers, torch, and sentence-transformers. It fetches pre-trained models from established providers like Google and OpenAI via the Hugging Face Hub, which is standard practice for machine learning development.
[INDIRECT_PROMPT_INJECTION]: The implementation examples for e-commerce and fashion search ingest user-provided metadata (such as category names or price ranges) and interpolate them into vector search filters using f-strings (e.g., filter=f'category == "{category}"'). While this is functional for demonstration purposes, it creates a surface for filter injection where a malicious input could potentially bypass intended search constraints.
Ingestion points: Search functions in verticals/ecommerce.md and verticals/fashion-search.md accept untrusted metadata strings for filtering.
Boundary markers: Absent; the metadata is directly embedded into the filter query string.
Capability inventory: The skill allows reading from Milvus collections and performing vector similarity searches.
Sanitization: The boilerplate code does not include input sanitization; developers should ensure inputs are validated or escaped before being used in filter expressions.

image-search