agent-contracts Tool Use / RAG

Use this skill when your nodes call external services (DB, search, HTTP APIs) or do retrieval-style work.

Core Ideas

• Declare dependencies in NodeContract.services (and requires_llm when needed).
• Inject dependencies via:
  • direct node construction (node_cls(llm=..., my_service=...)), or
  • dependency_provider in build_graph_from_registry(...).
• Keep LLM routing context small; rely on sanitization and context_builder only when needed.

Workflow

1. Define service interfaces (thin wrappers) and keep them testable.
2. Add services=[...] to NodeContract and use those attributes in execute().
3. Validate services with ContractValidator(known_services=..., strict=True) in CI.
4. For RAG:
   • store small retrieval results in a domain slice (ids + snippets)
   • avoid storing full documents or large embeddings in state
5. For routing accuracy:
   • use rule triggers for obvious cases
   • use LLM only among a small candidate set

Guardrails

• Sanitize long strings and binary-like content before LLM routing (supervisor does this).
• Avoid leaking secrets into state; treat state as loggable.

References (load only when needed)

• docs/core_concepts.md (Context Builder / Sanitization)
• docs/skills/official/agent-contracts-tool-use-rag/references/di_and_testing.md