The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The code in references/chunking-strategies.md uses the trust_remote_code=True parameter when loading the jinaai/jina-embeddings-v2-base-en model via the transformers library. This setting allows the execution of arbitrary Python code provided by the model author on HuggingFace, which can pose a significant security risk if the remote repository is compromised.
[PROMPT_INJECTION]: The skill processes external data (user queries, chat history, and document chunks) by interpolating them into LLM prompts, creating an indirect prompt injection surface.
Ingestion points: Untrusted data enters the context via user queries and document text processed in references/retrieval-optimization.md and references/chunking-strategies.md.
Boundary markers: The implementation examples lack explicit delimiters or instructions to ignore embedded commands within the processed data.
Capability inventory: The skill environment includes tools for network operations and database access.
Sanitization: There is no evidence of input validation or sanitization in the provided prompts.
[EXTERNAL_DOWNLOADS]: The skill's implementation scripts perform multiple remote requests to download resources. This includes fetching datasets using nltk.download('punkt') in references/chunking-strategies.md and downloading model weights from HuggingFace. While these sources are well-known, the reliance on external assets at runtime is a risk factor.

rag-architect