RAG Agent that calls external retrieval server

[xingdi-eric-yuan]

This pull request introduces a new RAGAgent (Retrieval-Augmented Generation Agent) to the debug_gym framework, enabling retrieval-based decision-making by integrating with a standalone retrieval service. The changes include the implementation of the agent, configuration updates, and mock-based tests for validation. Below is a summary of the most important changes:

New Feature: RAGAgent Implementation

• Added RAGAgent class in debug_gym/agents/rag_agent.py, which integrates with a retrieval service to retrieve relevant examples for decision-making. It includes configuration options for indexing methods, retrieval parameters, and caching. The agent conditionally imports the retrieval service client and raises an error if unavailable.
• Updated debug_gym/agents/__init__.py to conditionally import RAGAgent only if the retrieval service is installed, ensuring compatibility with setups without the retrieval service.

Configuration Updates

• Added a configuration section for rag_agent in scripts/config_swesmith.yaml, specifying tools, retrieval parameters, and the experience trajectory path for building indexes.

Code Refinements

• Enhanced the run_agent function in scripts/run.py to refine handling of previous run statuses, ensuring better control over task execution. [1] [2]

Testing Enhancements

• Introduced tests/agents/test_rag_agent_mock_only.py to provide mock-based tests for RAGAgent, covering import error handling, indexing method parsing, query text extraction, configuration defaults, retrieval service client interface, and prompt-building logic. These tests ensure the agent's logic can be validated without requiring the retrieval service.