The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill facilitates the creation of data transformation pipelines that ingest content from various external sources and process it using Large Language Models (LLMs), creating a potential surface for indirect prompt injection.
Ingestion points: The skill supports data ingestion from local files (LocalFile), Amazon S3 buckets (AmazonS3), and Postgres databases (Postgres), as described in SKILL.md and references/flow_patterns.md.
Boundary markers: There are no specific instructions or examples provided for implementing boundary markers or system prompt delimiters to isolate user-provided data from the transformation instructions in functions like ExtractByLlm.
Capability inventory: The skill possesses significant capabilities, including executing shell commands via the cocoindex CLI, writing to multiple database types (Postgres, Qdrant, Neo4j, LanceDB), and performing network operations to interact with LLM APIs and cloud storage.
Sanitization: The provided documentation and examples do not include mechanisms for sanitizing or validating external data before it is interpolated into LLM prompts for extraction or summarization tasks.

cocoindex