hybrid-search-implementation
Pass
Audited by Gen Agent Trust Hub on Sep 14, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [SAFE]: The skill contains educational code templates for Reciprocal Rank Fusion (RRF) and linear combination search strategies. It utilizes trusted libraries such as sentence-transformers for reranking search results using pre-trained models from Hugging Face. The implementation follows common patterns for modern search architecture.\n- [INDIRECT_PROMPT_INJECTION]: The skill provides a framework for processing external data from databases, creating a potential surface for indirect prompt injection if the retrieved content is used in downstream LLM prompts without sanitization.\n
- Ingestion points: Data is ingested via
hybrid_searchinPostgresHybridSearchandElasticsearchHybridSearch(SKILL.md).\n - Boundary markers: None identified in the provided templates.\n
- Capability inventory: The skill provides database querying and result ranking capabilities.\n
- Sanitization: No explicit sanitization of retrieved content is performed in the templates.\n- [COMMAND_EXECUTION]: Note: The PostgreSQL search template uses string interpolation for metadata keys in the
WHEREclause (metadata->>'{key}'), which could allow SQL injection if keys are provided by untrusted users. This is a common pattern in search templates but requires caution in production environments.
Audit Metadata