The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill and its reference guides utilize a wide range of official libraries for building AI pipelines, including langchain, openai, qdrant-client, rank_bm25, cohere, ragas, datasets, tiktoken, nltk, sentence-transformers, scikit-learn, numpy, transformers, torch, FlagEmbedding, weaviate-client, pinecone-client, pinecone-text, chromadb, psycopg2, trulens-eval, and tqdm. These are well-known packages from established organizations used for their primary intended purpose.
[INDIRECT_PROMPT_INJECTION]: The RAG architecture described processes external data from untrusted sources, which presents an inherent indirect prompt injection surface.
Ingestion points: The skill ingests raw documents (e.g., raw_docs, document_text, long_document) to create chunks and embeddings, and handles user queries at runtime for retrieval.
Boundary markers: The provided code snippets do not implement explicit security boundary markers or instructions to ignore embedded commands within retrieved context.
Capability inventory: The system is designed to perform network operations to AI service providers (OpenAI, Cohere) and vector databases (Qdrant, Pinecone, Weaviate).
Sanitization: While the code provides text cleaning for processing quality (e.g., regex-based whitespace removal), it does not include security-focused sanitization to prevent the injection of instructions into the final prompt.
[DYNAMIC_EXECUTION]: The LateChunker implementation in references/chunking-strategies.md uses the trust_remote_code=True flag when loading models from Hugging Face (jinaai/jina-embeddings-v2-base-en). This flag enables the execution of code defined in the model's repository. While the source is a recognized AI company, this pattern allows for remote code execution if the model repository were compromised or if the model name were user-controlled.

rag-architect