Improve search tool to extract resolved urls

[amrit110]

This pull request introduces a major overhaul to the Google Search tool for agentic evaluation, shifting from returning redirect URLs to providing actual, fetchable URLs for agent workflows. It adds robust redirect resolution utilities, updates the tool to use Gemini search with grounding, and ensures the agent can verify sources directly. The changes also update dependencies and tests to support the new workflow.

Search Tool Redesign and Redirect Resolution:

• Added a new module aieng/agent_evals/tools/_redirect.py providing sync and async utilities to resolve Vertex AI grounding redirect URLs to their final destinations, with caching, retries, and robust error handling.
• Refactored aieng/agent_evals/tools/search.py to implement a new search tool using Gemini with Google Search grounding, extracting actual URLs, resolving redirects, and returning structured results for agent verification. The legacy GoogleSearchTool usage was removed. [1] [2] [3]

Testing and Interface Updates:

• Updated tests in tests/aieng/agent_evals/tools/test_search.py to validate that the new search tool returns real URLs, is wrapped in a FunctionTool, and provides the correct response structure for agent workflows. [1] [2] [3]

Dependency and Version Updates:

• Updated aieng-eval-agents/pyproject.toml to bump the package version to 0.2.0 and add new dependencies: httpx, tenacity, python-dotenv, html-to-markdown, and pypdf for improved web and PDF handling. Also added a CLI script entry point. [1] [2]

Agent Evaluation Compatibility:

• Updated implementations/report_generation/evaluate.py to support both ExperimentItem and legacy dataset item types for agent evaluation, improving compatibility with Langfuse experiments. [1] [2] [3]

Minor Cleanup:

• Removed the unused weaviate-client dependency from the root pyproject.toml.