refactor(llm): share summary prompt builder

lib/crewai/src/crewai/llms/context_window.py

@@ -26,10 +26,10 @@ from crewai.utilities.exceptions.context_window_exceeding_exception import (
 )
 from crewai.utilities.i18n import I18N_DEFAULT
 from crewai.utilities.message_utils import (
+    build_summary_prompt,
     collect_user_message_files,
     format_message_for_llm,
     message_content_text,
-    messages_content_text,
 )
 
 
@@ -316,7 +316,7 @@ def _summarize_chunks(
         try:
             async with semaphore:
                 response = str(
-                    await llm.acall(_build_summary_prompt(chunk), callbacks=callbacks)
+                    await llm.acall(build_summary_prompt(chunk), callbacks=callbacks)
                 )
         except Exception as error:
             if not LLMContextLengthExceededError(
@@ -403,19 +403,6 @@ def _compact_summaries(
     )
 
 
-def _build_summary_prompt(chunk: list[LLMMessage]) -> list[LLMMessage]:
-    return [
-        format_message_for_llm(
-            I18N_DEFAULT.slice("summarizer_system_message"), "system"
-        ),
-        format_message_for_llm(
-            I18N_DEFAULT.slice("summarize_instruction").format(
-                conversation=messages_content_text(chunk)
-            )
-        ),
-    ]
-
-
 def _chunk_messages(
     llm: Any, messages: list[LLMMessage], max_tokens: int
 ) -> list[list[LLMMessage]]:
@@ -501,4 +488,4 @@ def _split_message_to_fit(
 
 def _summary_request_tokens(llm: Any, messages: list[LLMMessage]) -> int:
     """Count the exact request used for a summary call through the active LLM."""
-    return int(llm.get_input_tokens_count(_build_summary_prompt(messages)))
+    return int(llm.get_input_tokens_count(build_summary_prompt(messages)))

lib/crewai/src/crewai/utilities/message_utils.py

@@ -4,6 +4,7 @@ from __future__ import annotations
 
 from typing import Any, Literal
 
+from crewai.utilities.i18n import I18N_DEFAULT
 from crewai.utilities.types import LLMMessage
 
 
@@ -57,6 +58,20 @@ def messages_content_text(messages: list[LLMMessage]) -> str:
     return "\n\n".join(lines)
 
 
+def build_summary_prompt(chunk: list[LLMMessage]) -> list[LLMMessage]:
+    """Build the LLM request used to summarize a conversation chunk."""
+    return [
+        format_message_for_llm(
+            I18N_DEFAULT.slice("summarizer_system_message"), "system"
+        ),
+        format_message_for_llm(
+            I18N_DEFAULT.slice("summarize_instruction").format(
+                conversation=messages_content_text(chunk)
+            )
+        ),
+    ]
+
+
 def collect_user_message_files(messages: list[LLMMessage]) -> dict[str, Any]:
     """Merge files attached to user messages."""
     files: dict[str, Any] = {}

lib/crewai/tests/llms/test_context_window.py

@@ -13,11 +13,11 @@ from crewai.llms.context_window import (
     LLM_CONTEXT_WINDOW_SIZES,
     bare_context_windows,
     context_windows_from_catalog,
-    _build_summary_prompt,
     _chunk_messages,
     summarize_messages,
     resolve_context_window_size,
 )
+from crewai.utilities.message_utils import build_summary_prompt
 from crewai.utilities.exceptions.context_window_exceeding_exception import (
     LLMContextLengthExceededError,
 )
@@ -283,7 +283,7 @@ def test_summary_chunking_uses_the_rendered_request_token_estimator() -> None:
 
     assert chunks == [[messages[0]], [messages[1]]]
     assert all(
-        llm.get_input_tokens_count(_build_summary_prompt(chunk)) <= 125
+        llm.get_input_tokens_count(build_summary_prompt(chunk)) <= 125
         for chunk in chunks
     )
     assert all(len(request) == 2 for request in llm.requests)
@@ -411,7 +411,7 @@ def test_summary_chunking_splits_a_single_oversized_text_message() -> None:
     assert len(chunks) == 3
     assert "".join(chunk[0]["content"] for chunk in chunks) == message["content"]
     assert all(
-        CountingLLM().get_input_tokens_count(_build_summary_prompt(chunk)) <= 20
+        CountingLLM().get_input_tokens_count(build_summary_prompt(chunk)) <= 20
         for chunk in chunks
     )