Add LLM Evaluation Feature

[wanhakim]

Introduces a full evaluation suite for testing and benchmarking deployed GenAI workflows within the Studio.

What Is New

studio-eval microservice (new FastAPI service)

• Manage evaluation datasets: create manually via JSON or JSONL upload, or synthesize from documents (PDF, DOCX) using DeepEval and Ollama.
• Enabled evaluation runs against active sandboxes with configurable metrics, currently supporting Answer Relevancy and Faithfulness (with additional metrics easily extendable via DeepEval's framework).
• Support background job execution with real-time progress tracking (polling-based).
• Persist runs, datasets, entries, and per-entry metric results (SQLite-backed).
• Expose CRUD APIs for runs (/eval/runs) and datasets (/eval/datasets).

Frontend UI updates (studio-frontend)

• Add a new Evaluation section with Execution and Dataset tabs.
• Dataset tab supports upload, manual entry creation, synthesis configuration, and synthesis progress.
• Execution tab supports run creation (sandbox and judge model selection), progress tracking, and metric summaries.
• RunDetailsModal shows per-entry scores, reason drill-down, and configuration snapshot.
• Add ModelSelect UI to browse and pull Ollama judge models directly from the UI.

studio-backend proxy layer

• Add evaluation_router.py to proxy /evaluation/{path} requests to studio-eval.
• Extend sandbox APIs to support sandbox discovery for evaluation workflows.

Infrastructure Changes

• Development setup: update local compose and nginx config templates, and provide .env.example entries for evaluation stack.
• Kubernetes setup: update manifests and playbooks to deploy studio-eval and Ollama in cluster environments.
• CI pipeline: update image build and push workflow to include studio-eval artifacts.

Bug Fixes and Improvements

• Fix deprecated TGI API usage that broke existing workflows.
• Improve UX in tracer view under observability from workflow section.

[Copilot]

Copilot wasn't able to review this pull request because it exceeds the maximum number of lines (20,000). Try reducing the number of changed lines and requesting a review from Copilot again.

[siawchen]

lgtm