feat: Resolve Interactions API data for agent trace display in eval results.

PiperOrigin-RevId: 944658614

Author: vertex-sdk-bot

agentplatform/_genai/_evals_common.py

@@ -923,6 +923,125 @@ def _fetch_agent_config_dict(
     )
 
 
+def _resolve_interactions_for_display(
+    api_client: BaseApiClient,
+    dataset_list: list[types.EvaluationDataset],
+) -> list[types.EvaluationDataset]:
+    """Resolves interactions_data_source on EvalCases for display.
+
+    For each EvalCase that has ``interactions_data_source`` set but no
+    ``agent_data``, fetches the Interaction and Agent config from the
+    API, converts to AgentData, and populates it on the EvalCase so
+    that ``show()`` can render the System Topology and Conversation
+    Trace sections.
+
+    Args:
+        api_client: The API client used to fetch interactions and agents.
+        dataset_list: The original evaluation datasets.
+
+    Returns:
+        A list of EvaluationDatasets with agent_data populated from
+        resolved interactions.
+    """
+    resolved_datasets = []
+    for dataset in dataset_list:
+        if not dataset.eval_cases:
+            resolved_datasets.append(dataset)
+            continue
+
+        resolved_cases = []
+        any_resolved = False
+        for case in dataset.eval_cases:
+            ids = case.interactions_data_source
+            if ids and not case.agent_data:
+                interaction_name = ids.interaction
+                gemini_cfg = ids.gemini_agent_config
+                agent_name = gemini_cfg.gemini_agent if gemini_cfg else None
+
+                if interaction_name:
+                    try:
+                        # Extract the interaction short ID from the resource
+                        # name; handles both full resource names and bare IDs.
+                        interaction_id = interaction_name.split("/")[-1]
+
+                        logger.info(
+                            "Fetching interaction chain for %s for display.",
+                            interaction_name,
+                        )
+                        current_interaction_id = interaction_id
+                        interactions = []
+                        seen_ids = set()
+                        for _ in range(10): # max depth
+                            if current_interaction_id in seen_ids:
+                                break
+                            seen_ids.add(current_interaction_id)
+                            path = f"interactions/{current_interaction_id}"
+                            response = api_client.request("get", path, {}, None)
+                            if not response.body:
+                                if not interactions:
+                                    logger.warning(
+                                        "Empty response fetching interaction %s.",
+                                        interaction_name,
+                                    )
+                                break
+                            interaction_dict = json.loads(response.body)
+                            interactions.append(interaction_dict)
+                            prev_id_str = interaction_dict.get("previousInteractionId") or interaction_dict.get("previous_interaction_id")
+                            if not prev_id_str:
+                                break
+                            current_interaction_id = prev_id_str.split("/")[-1]
+
+                        if not interactions:
+                            resolved_cases.append(case)
+                            continue
+
+                        interactions.reverse() # chronological order
+                        all_steps = []
+                        for i_dict in interactions:
+                            all_steps.extend(i_dict.get("steps", []))
+
+                        combined_interaction = interactions[-1].copy()
+                        combined_interaction["steps"] = all_steps
+                        logger.warning(f"DEBUG: I fetched {len(interactions)} interactions. all_steps length: {len(all_steps)}")
+                        for i, idct in enumerate(interactions):
+                            logger.warning(f"DEBUG int {i}: {idct.get('id', 'no_id')} has {len(idct.get('steps', []))} steps")
+                        agent_data = _interaction_dict_to_agent_data(combined_interaction)
+
+                        # Best-effort: fetch agent config (instruction,
+                        # tools, description) from the Agent API.
+                        agent_config = _fetch_agent_config_dict(
+                            api_client, agent_name or ""
+                        )
+                        agent_data.agents = {agent_config.agent_id: agent_config}
+
+                        _merge_text_parts_in_agent_data(agent_data)
+
+                        # Preserve all original EvalCase fields; only
+                        # update agent_data.
+                        resolved_cases.append(
+                            case.model_copy(update={"agent_data": agent_data})
+                        )
+                        any_resolved = True
+                        continue
+                    except Exception as e:  # pylint: disable=broad-exception-caught
+                        logger.warning(
+                            "Failed to resolve interaction %s for display: %s",
+                            interaction_name,
+                            e,
+                        )
+
+            resolved_cases.append(case)
+
+        if any_resolved:
+            resolved_datasets.append(
+                dataset.model_copy(update={"eval_cases": resolved_cases})
+            )
+        else:
+            resolved_datasets.append(dataset)
+
+    return resolved_datasets
+
+
 def _add_evaluation_run_labels(
     labels: Optional[dict[str, str]] = None,
     agent: Optional[str] = None,
@@ -2245,6 +2364,11 @@ def _execute_evaluation(  # type: ignore[no-untyped-def]
     t2 = time.perf_counter()
     logger.info("Evaluation took: %f seconds", t2 - t1)
 
+    # Resolve interactions_data_source to agent_data for display.
+    # This fetches Interaction trace data client-side so that show() can
+    # render the System Topology and Conversation Trace sections.
+    dataset_list = _resolve_interactions_for_display(api_client, dataset_list)
+
     evaluation_result.evaluation_dataset = dataset_list
     evaluation_result.agent_info = validated_agent_info
 

tests/unit/agentplatform/genai/replays/test_evaluate.py

@@ -582,6 +582,57 @@ def test_evaluation_single_turn_agent_data(client):
     assert len(evaluation_result.eval_case_results) == 1
 
 
+def test_evaluation_with_interaction_resolution(client):
+    """Tests evaluate() resolves interaction trace + agent config for display."""
+    interaction_id = "ChAxMDE5YTgzOGU3Mjg4NDE4EAgaATAqBG1haW4"
+    agent = (
+        "projects/model-evaluation-dev/locations/global/agents/"
+        "async-scrape-test-fe111cc6"
+    )
+    eval_dataset = types.EvaluationDataset(
+        eval_cases=[
+            types.EvalCase(
+                interactions_data_source=types.InteractionsDataSource(
+                    interaction=(
+                        "projects/model-evaluation-dev/locations/global"
+                        f"/interactions/{interaction_id}"
+                    ),
+                    gemini_agent_config=types.GeminiAgentConfig(
+                        gemini_agent=agent,
+                    ),
+                ),
+            )
+        ]
+    )
+
+    evaluation_result = client.evals.evaluate(
+        dataset=eval_dataset,
+        metrics=[types.RubricMetric.MULTI_TURN_TASK_SUCCESS],
+    )
+
+    assert isinstance(evaluation_result, types.EvaluationResult)
+    assert evaluation_result.summary_metrics is not None
+    assert len(evaluation_result.summary_metrics) > 0
+
+    assert evaluation_result.eval_case_results is not None
+    assert len(evaluation_result.eval_case_results) == 1
+
+    # Verify the interaction was resolved for display: agent_data should
+    # be populated with conversation trace and agent topology.
+    eval_dataset_list = evaluation_result.evaluation_dataset
+    assert eval_dataset_list is not None
+    assert len(eval_dataset_list) > 0
+    resolved_case = eval_dataset_list[0].eval_cases[0]
+    assert resolved_case.agent_data is not None
+    # Conversation trace: at least one turn with events.
+    assert resolved_case.agent_data.turns is not None
+    assert len(resolved_case.agent_data.turns) > 0
+    assert len(resolved_case.agent_data.turns[0].events) > 0
+    # Agent topology: agents map populated with agent config.
+    assert resolved_case.agent_data.agents is not None
+    assert len(resolved_case.agent_data.agents) > 0
+
+
 pytestmark = pytest_helper.setup(
     file=__file__,
     globals_for_file=globals(),