From 82b8e304ecffc87ca011eadfd4098b9eacd7e887 Mon Sep 17 00:00:00 2001
From: Ankur Sharma <ankusharma@google.com>
Date: Tue, 13 May 2025 13:39:17 -0700
Subject: [PATCH] Convert Session To Invocation.

PiperOrigin-RevId: 758356942
---
 src/google/adk/cli/utils/evals.py      | 84 +++++++++++++++++++++++++-
 src/google/adk/evaluation/eval_case.py | 13 +++-
 2 files changed, 93 insertions(+), 4 deletions(-)

diff --git a/src/google/adk/cli/utils/evals.py b/src/google/adk/cli/utils/evals.py
index 2be9fb8..fb3ddf1 100644
--- a/src/google/adk/cli/utils/evals.py
+++ b/src/google/adk/cli/utils/evals.py
@@ -12,11 +12,17 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from typing import Any
+from typing import Any, Tuple
 
+from deprecated import deprecated
+from google.genai import types as genai_types
+
+from ...evaluation.eval_case import IntermediateData
+from ...evaluation.eval_case import Invocation
 from ...sessions.session import Session
 
 
+@deprecated(reason='Use convert_session_to_eval_invocations instead.')
 def convert_session_to_eval_format(session: Session) -> list[dict[str, Any]]:
   """Converts a session data into eval format.
 
@@ -91,3 +97,79 @@ def convert_session_to_eval_format(session: Session) -> list[dict[str, Any]]:
       })
 
   return eval_case
+
+
+def convert_session_to_eval_invocations(session: Session) -> list[Invocation]:
+  """Converts a session data into a list of Invocation.
+
+  Args:
+      session: The session that should be converted.
+
+  Returns:
+      list: A list of invocation.
+  """
+  invocations: list[Invocation] = []
+  events = session.events if session and session.events else []
+
+  for event in events:
+    if event.author == 'user':
+      if not event.content or not event.content.parts:
+        continue
+
+      # The content present in this event is the user content.
+      user_content = event.content
+      invocation_id = event.invocation_id
+      invocaton_timestamp = event.timestamp
+
+      # Find the corresponding tool usage or response for the query
+      tool_uses: list[genai_types.FunctionCall] = []
+      intermediate_responses: list[Tuple[str, list[genai_types.Part]]] = []
+
+      # Check subsequent events to extract tool uses or responses for this turn.
+      for subsequent_event in events[events.index(event) + 1 :]:
+        event_author = subsequent_event.author or 'agent'
+        if event_author == 'user':
+          # We found an event where the author was the user. This means that a
+          # new turn has started. So close this turn here.
+          break
+
+        if not subsequent_event.content or not subsequent_event.content.parts:
+          continue
+
+        intermediate_response_parts = []
+        for subsequent_part in subsequent_event.content.parts:
+          # Some events have both function call and reference
+          if subsequent_part.function_call:
+            tool_uses.append(subsequent_part.function_call)
+          elif subsequent_part.text:
+            # Also keep track of all the natural language responses that
+            # agent (or sub agents) generated.
+            intermediate_response_parts.append(subsequent_part)
+
+        if intermediate_response_parts:
+          # Only add an entry if there any intermediate entries.
+          intermediate_responses.append(
+              (event_author, intermediate_response_parts)
+          )
+
+      # If we are here then either we are done reading all the events or we
+      # encountered an event that had content authored by the end-user.
+      # This, basically means an end of turn.
+      # We assume that the last natural language intermediate response is the
+      # final response from the agent/model. We treat that as a reference.
+      invocations.append(
+          Invocation(
+              user_content=user_content,
+              invocation_id=invocation_id,
+              creation_timestamp=invocaton_timestamp,
+              intermediate_data=IntermediateData(
+                  tool_uses=tool_uses,
+                  intermediate_responses=intermediate_responses[:-1],
+              ),
+              final_response=genai_types.Content(
+                  parts=intermediate_responses[-1][1]
+              ),
+          )
+      )
+
+  return invocations
diff --git a/src/google/adk/evaluation/eval_case.py b/src/google/adk/evaluation/eval_case.py
index 9966a2d..16d07dd 100644
--- a/src/google/adk/evaluation/eval_case.py
+++ b/src/google/adk/evaluation/eval_case.py
@@ -13,7 +13,7 @@
 # limitations under the License.
 
 
-from typing import Any, Optional
+from typing import Any, Optional, Tuple
 
 from google.genai import types as genai_types
 from pydantic import BaseModel
@@ -26,9 +26,16 @@ class IntermediateData(BaseModel):
   tool_uses: list[genai_types.FunctionCall]
   """Tool use trajectory in chronological order."""
 
-  intermediate_responses: list[genai_types.Part]
+  intermediate_responses: list[Tuple[str, list[genai_types.Part]]]
   """Intermediate responses generated by sub-agents to convey progress or status
-  in a multi-agent system, distinct from the final response."""
+  in a multi-agent system, distinct from the final response.
+
+  This is expressed as a Tuple of:
+    - Author: Usually the sub-agent name that generated the intermediate
+      response.
+
+    - A list of Parts that comprise of the response.
+  """
 
 
 class Invocation(BaseModel):