The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill's chunking reference (references/chunking-strategies.md) includes an implementation of 'Late Chunking' that utilizes the transformers.AutoModel.from_pretrained method with the trust_remote_code=True parameter. This setting allows the execution of arbitrary Python code defined within the remote model repository. While the example targets a well-known model repository (jinaai/jina-embeddings-v2-base-en), this pattern introduces a risk of remote code execution if the repository is compromised or points to a malicious source.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process external, untrusted documents to build knowledge bases, creating a surface for indirect prompt injection.
Ingestion points: Multiple functions in SKILL.md and references/chunking-strategies.md ingest raw text from documents (doc.page_content).
Boundary markers: The LLM prompts used for contextual chunking and proposition extraction lack explicit delimiters or instructions to treat the ingested text as untrusted data.
Capability inventory: The skill possesses capabilities for network communication (multiple LLM and Vector DB APIs) and file system interactions (caching embeddings).
Sanitization: Documented preprocessing focuses on formatting and length rather than sanitizing against malicious instructions embedded in documents.
[EXTERNAL_DOWNLOADS]: The skill includes instructions to download external resources from well-known and trusted providers during runtime.
It fetches the punkt tokenizer data from NLTK's official servers in references/chunking-strategies.md.
It downloads model weights from Hugging Face for various embedding and evaluation tasks (sentence-transformers, FlagEmbedding).

rag-architect