The Agent Skills Directory

Indirect Prompt Injection: The skill facilitates the ingestion and processing of external document data for use in LLM grounding prompts, creating a potential surface for indirect injection via processed content.
Ingestion points: Document text is ingested into the dbo.documents and dbo.document_chunks tables as described in SKILL.md.
Boundary markers: The instructions emphasize incorporating a tenant filter (WHERE c.tenant_id = @tenant_id) directly into the VECTOR_SEARCH query in SKILL.md to maintain data boundaries, noting that traditional row-level security policies are restricted on vector-indexed tables.
Capability inventory: The skill utilizes SQL capabilities for data management and retrieval, including AI_GENERATE_EMBEDDINGS, MERGE for idempotent updates, and VECTOR_SEARCH for similarity analysis, as detailed in both SKILL.md and references/pipeline.md.
Sanitization: The skill provides grounding instructions in SKILL.md that recommend source citations and explicit instructions for the model to decline answering if the retrieved context is insufficient.
Command Execution: The skill includes instructions for using the sqlcmd utility to perform non-interactive integrity checks on the database. These checks are used to verify embedding provenance, index versions, and the effectiveness of the multi-tenant filtering logic, ensuring the pipeline operates as intended within the Azure SQL environment.

rag-on-azure-sql