unique_toolkit 1.35.1__tar.gz → 1.35.3__tar.gz

{unique_toolkit-1.35.1 → unique_toolkit-1.35.3}/CHANGELOG.md

@@ -5,6 +5,12 @@ 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).
 
+## [1.35.3] - 2025-12-10
+- Add option to ignore some options when calling the LLM for the planning step.
+
+## [1.35.2] - 2025-12-05
+- Increase speed of token reducer
+
 ## [1.35.1] - 2025-12-05
 - Improve efficiency of token reducer if tool calls overshoot max token limit
 

{unique_toolkit-1.35.1 → unique_toolkit-1.35.3}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: unique_toolkit
-Version: 1.35.1
+Version: 1.35.3
 Summary: 
 License: Proprietary
 Author: Cedric Klinkert
@@ -121,6 +121,12 @@ 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).
 
+## [1.35.3] - 2025-12-10
+- Add option to ignore some options when calling the LLM for the planning step.
+
+## [1.35.2] - 2025-12-05
+- Increase speed of token reducer
+
 ## [1.35.1] - 2025-12-05
 - Improve efficiency of token reducer if tool calls overshoot max token limit
 

{unique_toolkit-1.35.1 → unique_toolkit-1.35.3}/pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "unique_toolkit"
-version = "1.35.1"
+version = "1.35.3"
 description = ""
 authors = [
     "Cedric Klinkert <cedric.klinkert@unique.ch>",

{unique_toolkit-1.35.1 → unique_toolkit-1.35.3}/unique_toolkit/agentic/history_manager/loop_token_reducer.py

@@ -31,6 +31,7 @@ from unique_toolkit.language_model.schemas import (
 MAX_INPUT_TOKENS_SAFETY_PERCENTAGE = (
     0.1  # 10% safety margin for input tokens we need 10% less does not work.
 )
+SAFETY_MARGIN_FOR_AGGRESSIVE_REDUCTION = 0.9
 
 
 class SourceReductionResult(BaseModel):
@@ -323,10 +324,10 @@ class LoopTokenReducer:
         Reduce the message length by removing sources from each tool call based on overshoot.
 
         The number of chunks to keep per tool call is calculated as:
-        chunks_to_keep = num_sources / (overshoot_factor * 0.75)
+        chunks_to_keep = num_sources / (overshoot_factor * SAFETY_MARGIN_FOR_AGGRESSIVE_REDUCTION)
 
         This ensures more aggressive reduction when we're significantly over the limit.
-        Using 0.75 factor provides a safety margin to avoid over-reduction.
+        Using SAFETY_MARGIN_FOR_AGGRESSIVE_REDUCTION factor provides a safety margin to avoid over-reduction.
         E.g., if overshoot_factor = 2 (2x over limit), keep 1/1.5 = 2/3 of chunks.
         Always keeps at least 1 chunk.
         """
@@ -369,7 +370,7 @@ class LoopTokenReducer:
         """
         Reduce the sources in the tool message based on overshoot factor.
 
-        Chunks to keep = num_sources / (overshoot_factor * 0.75)
+        Chunks to keep = num_sources / (overshoot_factor * SAFETY_MARGIN_FOR_AGGRESSIVE_REDUCTION)
         This ensures fewer chunks are kept when overshoot is larger.
         E.g., if overshoot_factor = 2 (2x over limit), keep 1/1.5 = 2/3 of chunks
         Always keeps at least 1 chunk.
@@ -386,14 +387,15 @@ class LoopTokenReducer:
             )
 
         # Calculate how many chunks to keep based on overshoot
-        # Use 0.75 safety margin for aggressive reduction, but only when overshoot is
+        # Use SAFETY_MARGIN_FOR_AGGRESSIVE_REDUCTION safety margin for aggressive reduction, but only when overshoot is
         # significant enough (>= ~1.33). Otherwise, the margin would prevent reduction.
         divisor = (
-            overshoot_factor * 0.75
-            if overshoot_factor * 0.75 >= 1.0
+            overshoot_factor * SAFETY_MARGIN_FOR_AGGRESSIVE_REDUCTION
+            if overshoot_factor * SAFETY_MARGIN_FOR_AGGRESSIVE_REDUCTION
+            >= 1.2  # Skip safety margin for small overshoots (e.g., 1.03) to avoid tiny reduction steps
             else overshoot_factor
         )
-        chunks_to_keep = max(1, int(num_sources / divisor))
+        chunks_to_keep = max(1, int(num_sources // divisor))
 
         # Reduce chunks
         if chunks_to_keep >= num_sources:

{unique_toolkit-1.35.1 → unique_toolkit-1.35.3}/unique_toolkit/agentic/loop_runner/middleware/planning/planning.py

@@ -2,7 +2,7 @@ import json
 import logging
 from typing import Unpack
 
-from pydantic import BaseModel
+from pydantic import BaseModel, Field
 
 from unique_toolkit import LanguageModelService
 from unique_toolkit._common.pydantic_helpers import get_configuration_dict
@@ -28,6 +28,10 @@ class PlanningConfig(BaseModel):
     model_config = get_configuration_dict()
 
     planning_schema_config: PlanningSchemaConfig = PlanningSchemaConfig()
+    ignored_options: list[str] = Field(
+        default=["parallel_tool_calls"],
+        description="A list of options to ignore when calling the LLM for the planning step.",
+    )
 
 
 class PlanningMiddleware(LoopIterationRunner):
@@ -50,11 +54,16 @@ class PlanningMiddleware(LoopIterationRunner):
     ) -> LanguageModelAssistantMessage | None:
         planning_schema = get_planning_schema(self._config.planning_schema_config)
 
+        other_options = {
+            k: v
+            for k, v in kwargs.get("other_options", {}).items()
+            if k not in self._config.ignored_options
+        }
         response = await self._llm_service.complete_async(
             messages=kwargs["messages"],
             model_name=kwargs["model"].name,
             structured_output_model=planning_schema,
-            other_options=kwargs.get("other_options", {}),
+            other_options=other_options,
         )
 
         if response.choices[0].message.parsed is None: