unique_toolkit 0.8.48__py3-none-any.whl → 0.8.50__py3-none-any.whl

unique_toolkit/history_manager/loop_token_reducer.py

@@ -28,6 +28,10 @@ from unique_toolkit.language_model.schemas import (
 )
 from unique_toolkit.reference_manager.reference_manager import ReferenceManager
 
+MAX_INPUT_TOKENS_SAFETY_PERCENTAGE = (
+    0.1  # 10% safety margin for input tokens we need 10% less does not work.
+)
+
 
 class SourceReductionResult(BaseModel):
     message: LanguageModelToolMessage
@@ -74,26 +78,27 @@ class LoopTokenReducer:
     ) -> LanguageModelMessages:
         """Compose the system and user messages for the plan execution step, which is evaluating if any further tool calls are required."""
 
-        messages = await self._construct_history(
+        history_from_db = await self._prep_db_history(
             original_user_message,
             rendered_user_message_string,
             rendered_system_message_string,
-            loop_history,
             remove_from_text,
         )
 
+        messages = self._construct_history(
+            history_from_db,
+            loop_history,
+        )
+
         token_count = self._count_message_tokens(messages)
         self._log_token_usage(token_count)
 
         while self._exceeds_token_limit(token_count):
             token_count_before_reduction = token_count
             loop_history = self._handle_token_limit_exceeded(loop_history)
-            messages = await self._construct_history(
-                original_user_message,
-                rendered_user_message_string,
-                rendered_system_message_string,
+            messages = self._construct_history(
+                history_from_db,
                 loop_history,
-                remove_from_text,
             )
             token_count = self._count_message_tokens(messages)
             self._log_token_usage(token_count)
@@ -101,6 +106,11 @@ class LoopTokenReducer:
             if token_count_after_reduction >= token_count_before_reduction:
                 break
 
+        token_count = self._count_message_tokens(messages)
+        self._logger.info(
+            f"Final token count after reduction: {token_count} of model_capacity {self._language_model.token_limits.token_limit_input}"
+        )
+
         return messages
 
     def _exceeds_token_limit(self, token_count: int) -> bool:
@@ -110,13 +120,14 @@ class LoopTokenReducer:
             len(chunks) > 1
             for chunks in self._reference_manager.get_chunks_of_all_tools()
         )
-
+        max_tokens = int(
+            self._language_model.token_limits.token_limit_input
+            * (1 - MAX_INPUT_TOKENS_SAFETY_PERCENTAGE)
+        )
         # TODO: This is not fully correct at the moment as the token_count
         # include system_prompt and user question already
         # TODO: There is a problem if we exceed but only have one chunk per tool call
-        exceeds_limit = (
-            token_count > self._language_model.token_limits.token_limit_input
-        )
+        exceeds_limit = token_count > max_tokens
 
         return has_multiple_chunks_for_a_tool_call and exceeds_limit
 
@@ -132,14 +143,13 @@ class LoopTokenReducer:
         self._logger.info(f"Token messages: {token_count}")
         # self.agent_debug_info.add("token_messages", token_count)
 
-    async def _construct_history(
+    async def _prep_db_history(
         self,
         original_user_message: str,
         rendered_user_message_string: str,
         rendered_system_message_string: str,
-        loop_history: list[LanguageModelMessage],
         remove_from_text: Callable[[str], Awaitable[str]],
-    ) -> LanguageModelMessages:
+    ) -> list[LanguageModelMessage]:
         history_from_db = await self._get_history_from_db(remove_from_text)
         history_from_db = self._replace_user_message(
             history_from_db, original_user_message, rendered_user_message_string
@@ -147,9 +157,15 @@ class LoopTokenReducer:
         system_message = LanguageModelSystemMessage(
             content=rendered_system_message_string
         )
+        return [system_message] + history_from_db
 
+    def _construct_history(
+        self,
+        history_from_db: list[LanguageModelMessage],
+        loop_history: list[LanguageModelMessage],
+    ) -> LanguageModelMessages:
         constructed_history = LanguageModelMessages(
-            [system_message] + history_from_db + loop_history,
+            history_from_db + loop_history,
         )
 
         return constructed_history

{unique_toolkit-0.8.48.dist-info → unique_toolkit-0.8.50.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: unique_toolkit
-Version: 0.8.48
+Version: 0.8.50
 Summary: 
 License: Proprietary
 Author: Cedric Klinkert
@@ -117,6 +117,14 @@ All notable changes to this project will be documented in this file.
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/), 
 and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
+
+## [0.8.50] - 2025-09-08
+- Minor fix in documentation
+- Updated examples
+
+## [0.8.49] - 2025-09-05
+- Fixed token reducer now has a safety margin of 10% less did not work.
+
 ## [0.8.48] - 2025-09-05
 - Add documentation on language models to markdown
 

{unique_toolkit-0.8.48.dist-info → unique_toolkit-0.8.50.dist-info}/RECORD

@@ -71,7 +71,7 @@ unique_toolkit/framework_utilities/openai/message_builder.py,sha256=VU6mJm_upLca
 unique_toolkit/framework_utilities/utils.py,sha256=JK7g2yMfEx3eMprug26769xqNpS5WJcizf8n2zWMBng,789
 unique_toolkit/history_manager/history_construction_with_contents.py,sha256=c8Zy3erSbHGT8AdICRRlSK91T_FN6tNpTznvUzpLbWk,9023
 unique_toolkit/history_manager/history_manager.py,sha256=7mdT8li4Oo-t0d1q0pCwJksdal-Y0wLzZj-YnIlJ6xQ,8350
-unique_toolkit/history_manager/loop_token_reducer.py,sha256=9kqJioUehfoYs5-XMoCv-b_5JpNRrhOz62QwhC3LF3E,17899
+unique_toolkit/history_manager/loop_token_reducer.py,sha256=YOWgEnodXlBeQMU3rA_wc6_qFHC_1ianuAvXN5ZZez4,18451
 unique_toolkit/history_manager/utils.py,sha256=iu4LsYOElx8HlZjcx3ZC75I-TmEYBiEP9q2J93Q63Mg,5606
 unique_toolkit/language_model/__init__.py,sha256=lRQyLlbwHbNFf4-0foBU13UGb09lwEeodbVsfsSgaCk,1971
 unique_toolkit/language_model/builder.py,sha256=4OKfwJfj3TrgO1ezc_ewIue6W7BCQ2ZYQXUckWVPPTA,3369
@@ -118,7 +118,7 @@ unique_toolkit/tools/utils/execution/execution.py,sha256=vjG2Y6awsGNtlvyQAGCTthQ
 unique_toolkit/tools/utils/source_handling/schema.py,sha256=vzAyf6ZWNexjMO0OrnB8y2glGkvAilmGGQXd6zcDaKw,870
 unique_toolkit/tools/utils/source_handling/source_formatting.py,sha256=C7uayNbdkNVJdEARA5CENnHtNY1SU6etlaqbgHNyxaQ,9152
 unique_toolkit/tools/utils/source_handling/tests/test_source_formatting.py,sha256=oM5ZxEgzROrnX1229KViCAFjRxl9wCTzWZoinYSHleM,6979
-unique_toolkit-0.8.48.dist-info/LICENSE,sha256=GlN8wHNdh53xwOPg44URnwag6TEolCjoq3YD_KrWgss,193
-unique_toolkit-0.8.48.dist-info/METADATA,sha256=JJiemqChaHLXodVnd3Cj7DNXMURuyF4--7H1p8oe-gA,31005
-unique_toolkit-0.8.48.dist-info/WHEEL,sha256=sP946D7jFCHeNz5Iq4fL4Lu-PrWrFsgfLXbbkciIZwg,88
-unique_toolkit-0.8.48.dist-info/RECORD,,
+unique_toolkit-0.8.50.dist-info/LICENSE,sha256=GlN8wHNdh53xwOPg44URnwag6TEolCjoq3YD_KrWgss,193
+unique_toolkit-0.8.50.dist-info/METADATA,sha256=8YBM58kjR4VR74qWcSLLCbh7avpqoUmwNF3UY8A5ZdI,31178
+unique_toolkit-0.8.50.dist-info/WHEEL,sha256=sP946D7jFCHeNz5Iq4fL4Lu-PrWrFsgfLXbbkciIZwg,88
+unique_toolkit-0.8.50.dist-info/RECORD,,