The Agent Skills Directory

[PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface common in RAG architectures. Ingestion points: Document processing logic in references/chunking-strategies.md and references/retrieval-optimization.md accepts external text data. Boundary markers: LLM prompt templates (e.g., in contextual_chunk and compress_retrieved_context) lack robust delimiters or specific instructions to ignore embedded commands within retrieved text. Capability inventory: Code samples demonstrate sending retrieved document content directly to LLM providers like OpenAI for summarization. Sanitization: No explicit sanitization or filtering of input text is performed before interpolation into LLM prompts.
[EXTERNAL_DOWNLOADS]: The reference code includes patterns for downloading resources from well-known and trusted technology providers. Downloads NLTK tokenizer data via nltk.download('punkt'). Fetches pre-trained embedding models from Hugging Face repositories such as BAAI/bge-large-en-v1.5 and jinaai/jina-embeddings-v2-base-en. Utilizes official SDKs for OpenAI, Cohere, Voyage AI, Pinecone, Weaviate, and Qdrant.
[REMOTE_CODE_EXECUTION]: Inclusion of trust_remote_code=True in model loading logic. Evidence: The LateChunker class in references/chunking-strategies.md demonstrates loading a Jina AI embedding model from Hugging Face with remote code execution enabled. While this is a common requirement for specialized model architectures from well-known vendors, it represents a theoretical execution vector for code included in the model repository.

rag-architect