[rag evals][3/n] add ability to eval retrieval + generation in agentic eval pipeline (#668)

# What does this PR do?

- This PR adds the ability s.t. users can evaluate on both retrieval +
generation separately & as a whole by passing an AgentConfig to the
/eval API
- The memory_retrieval context will be stored in the "context" column
used for scoring functions that can evaluate the retrieved context.

## Test Plan
- E2E Test RAG Agent Notebook:
https://gist.github.com/yanxi0830/0377594d29958f9b6f9317ab049fa836

<img width="758" alt="image"
src="https://github.com/user-attachments/assets/58ed9db7-f07b-400a-931b-923b0d612902"
/>

<img width="682" alt="image"
src="https://github.com/user-attachments/assets/9ebd7fbd-2a6d-4c93-92fa-a9456fae2378"
/>



## Sources

Please link relevant resources if necessary.


## Before submitting

- [ ] This PR fixes a typo or improves the docs (you can dismiss the
other checks if that's the case).
- [ ] Ran pre-commit to handle lint / formatting issues.
- [ ] Read the [contributor
guideline](https://github.com/meta-llama/llama-stack/blob/main/CONTRIBUTING.md),
      Pull Request section?
- [ ] Updated relevant documentation.
- [ ] Wrote necessary unit or integration tests.

llama_stack/providers/inline/eval/meta_reference/eval.py

@@ -7,7 +7,7 @@ from typing import Any, Dict, List, Optional
 
 from tqdm import tqdm
 
-from llama_stack.apis.agents import Agents
+from llama_stack.apis.agents import Agents, StepType
 from llama_stack.apis.datasetio import DatasetIO
 from llama_stack.apis.datasets import Datasets
 from llama_stack.apis.eval_tasks import EvalTask
@@ -139,11 +139,21 @@ class MetaReferenceEvalImpl(Eval, EvalTasksProtocolPrivate, DataSchemaValidatorM
                 )
             ]
             final_event = turn_response[-1].event.payload
-            generations.append(
+
-                {
+            # check if there's a memory retrieval step and extract the context
-                    ColumnName.generated_answer.value: final_event.turn.output_message.content
+            memory_rag_context = None
-                }
+            for step in final_event.turn.steps:
+                if step.step_type == StepType.memory_retrieval.value:
+                    memory_rag_context = " ".join(x.text for x in step.inserted_context)
+
+            agent_generation = {}
+            agent_generation[ColumnName.generated_answer.value] = (
+                final_event.turn.output_message.content
             )
+            if memory_rag_context:
+                agent_generation[ColumnName.context.value] = memory_rag_context
+
+            generations.append(agent_generation)
 
         return generations