The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill documents the use of aiDocuments() to ingest content from various sources, including local directories and external URLs, which are then used as context for an AI agent.
Ingestion points: aiDocuments() is used in SKILL.md to load data from /path/to/docs, expandPath( "./docs" ), and the URL https://boxlang.ortusbooks.com.
Capability inventory: The retrieved content is stored in vectorMemory and subsequently processed by aiAgent or aiChat to generate responses.
Boundary markers: The skill provides an example of boundary instructions: "Answer questions using only the provided documentation. If unsure, say so." in the aiAgent definition. While helpful, these markers do not fully prevent adversarial content within the ingested documents from influencing the agent's behavior.
Sanitization: The provided code snippets do not demonstrate explicit sanitization or filtering of the content loaded via aiDocuments() before it is indexed or processed by the LLM.

bx-ai-rag