similarity-search-patterns
Pass
Audited by Gen Agent Trust Hub on Oct 27, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill is designed to build Retrieval-Augmented Generation (RAG) systems, which involve ingesting untrusted external data into the agent's context.
- Ingestion points: The
upsertandinsertmethods in the provided Python templates (Pinecone, Qdrant, pgvector, Weaviate) accept document content and metadata from external sources. - Boundary markers: The documentation explicitly instructs developers to derive tenant/ACL filters from authenticated context and warns that user-supplied filters are not a substitute for proper authorization.
- Capability inventory: The templates perform network operations to communicate with official vector database APIs. They do not involve arbitrary shell execution or file system writes.
- Sanitization: The pgvector template utilizes parameterized queries (via
asyncpg) to prevent SQL injection. Other templates utilize standard client SDKs. - [REMOTE_CODE_EXECUTION]: The templates reference well-known machine learning and database libraries (e.g.,
sentence-transformers,pinecone-client). These are standard tools for the described use case, and the skill does not attempt to download or execute unverified scripts from the internet.
Audit Metadata