unique_toolkit 1.46.2__tar.gz → 1.46.4__tar.gz

{unique_toolkit-1.46.2 → unique_toolkit-1.46.4}/CHANGELOG.md

@@ -5,6 +5,18 @@ 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).
 
+
+## [1.46.4] - 2026-02-10
+- Fix: Add `union_mode="left_to_right"` to `encoder_name` field so Pydantic tries `EncoderName` enum before `str`, enabling bundled tokenizers when UI sends string values
+
+## [1.46.3] - 2026-02-09
+### Fixed
+- Fix hallucination check crash when responses contain invalid source indices (e.g., from code blocks with array indexing)
+- Use existing `context_text_from_stream_response()` utility with built-in bounds checking instead of manual extraction
+
+### Removed
+- Remove unused `regex` dependency (no longer needed after refactoring)
+
 ## [1.46.2] - 2026-02-09
 - Add `litellm:anthropic-claude-opus-4-6` and `AZURE_MODEL_ROUTER_2025_1118` to `info.py`
 

{unique_toolkit-1.46.2 → unique_toolkit-1.46.4}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: unique_toolkit
-Version: 1.46.2
+Version: 1.46.4
 Summary: 
 License: Proprietary
 Author: Cedric Klinkert
@@ -24,7 +24,6 @@ Requires-Dist: pydantic (>=2.8.2,<3.0.0)
 Requires-Dist: pydantic-settings (>=2.10.1,<3.0.0)
 Requires-Dist: pyhumps (>=3.8.0,<4.0.0)
 Requires-Dist: python-docx (>=1.1.0,<2.0.0)
-Requires-Dist: regex (>=2024.5.15,<2025.0.0)
 Requires-Dist: requests (>=2.32.0,<3.0.0)
 Requires-Dist: sseclient (>=0.0.27,<0.0.28)
 Requires-Dist: tiktoken (>=0.12.0,<0.13.0)
@@ -126,6 +125,18 @@ 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).
 
+
+## [1.46.4] - 2026-02-10
+- Fix: Add `union_mode="left_to_right"` to `encoder_name` field so Pydantic tries `EncoderName` enum before `str`, enabling bundled tokenizers when UI sends string values
+
+## [1.46.3] - 2026-02-09
+### Fixed
+- Fix hallucination check crash when responses contain invalid source indices (e.g., from code blocks with array indexing)
+- Use existing `context_text_from_stream_response()` utility with built-in bounds checking instead of manual extraction
+
+### Removed
+- Remove unused `regex` dependency (no longer needed after refactoring)
+
 ## [1.46.2] - 2026-02-09
 - Add `litellm:anthropic-claude-opus-4-6` and `AZURE_MODEL_ROUTER_2025_1118` to `info.py`
 

{unique_toolkit-1.46.2 → unique_toolkit-1.46.4}/pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "unique_toolkit"
-version = "1.46.2"
+version = "1.46.4"
 description = ""
 authors = [
     "Cedric Klinkert <cedric.klinkert@unique.ch>",
@@ -21,7 +21,6 @@ typing-extensions = "^4.9.0"
 pydantic = "^2.8.2"
 pyhumps = "^3.8.0"
 numpy = "^2.4.2"
-regex = "^2024.5.15"
 tiktoken = "^0.12.0"
 tokenizers = "^0.21.0"
 pydantic-settings = "^2.10.1"

{unique_toolkit-1.46.2 → unique_toolkit-1.46.4}/unique_toolkit/agentic/evaluation/hallucination/hallucination_evaluation.py

@@ -1,10 +1,11 @@
-import regex as re
-
 from unique_toolkit.agentic.evaluation.evaluation_manager import Evaluation
 from unique_toolkit.agentic.evaluation.hallucination.constants import (
     HallucinationConfig,
 )
-from unique_toolkit.agentic.evaluation.hallucination.utils import check_hallucination
+from unique_toolkit.agentic.evaluation.hallucination.utils import (
+    check_hallucination,
+    context_text_from_stream_response,
+)
 from unique_toolkit.agentic.evaluation.schemas import (
     EvaluationAssessmentMessage,
     EvaluationMetricInput,
@@ -20,7 +21,6 @@ from unique_toolkit.chat.schemas import (
     ChatMessageAssessmentStatus,
     ChatMessageAssessmentType,
 )
-from unique_toolkit.language_model.reference import _preprocess_message
 from unique_toolkit.language_model.schemas import (
     LanguageModelStreamResponse,
 )
@@ -45,24 +45,21 @@ class HallucinationEvaluation(Evaluation):
     ) -> EvaluationMetricResult:  # type: ignore
         all_chunks = self._reference_manager.get_chunks()
 
-        # source numbers from original text
-        ref_pattern = r"\[(\d+)\]"
-        original_text = loop_response.message.original_text
-
-        # preprocess original text to deal with different source patterns
-        original_text_preprocessed = _preprocess_message(original_text)
-
-        source_number_matches = re.findall(ref_pattern, original_text_preprocessed)
-        source_numbers = {int(num) for num in source_number_matches}
-
-        referenced_chunks = [all_chunks[idx] for idx in source_numbers]
+        # Extract context texts using existing utility function with bounds checking
+        # This prevents IndexError from invalid source indices (e.g., from code blocks)
+        context_texts = context_text_from_stream_response(
+            response=loop_response,
+            selected_chunks=all_chunks,
+            source_selection_mode=self.config.source_selection_mode,
+            reference_pattern=self.config.reference_pattern,
+        )
 
         evaluation_result: EvaluationMetricResult = await check_hallucination(
             company_id=self._company_id,
             user_id=self._user_id,
             input=EvaluationMetricInput(
                 input_text=self._user_message,
-                context_texts=[context.text for context in referenced_chunks],
+                context_texts=context_texts,
                 history_messages=[],  # TODO include loop_history messages
                 output_text=loop_response.message.text,
             ),

{unique_toolkit-1.46.2 → unique_toolkit-1.46.4}/unique_toolkit/language_model/infos.py

@@ -220,7 +220,7 @@ class LanguageModelInfo(BaseModel):
     version: str = Field(title="Model Version", default="")
 
     encoder_name: EncoderName | Annotated[str, Field(title="Custom Encoder Name")] = (
-        EncoderName.CL100K_BASE
+        Field(default=EncoderName.CL100K_BASE, union_mode="left_to_right")
     )
 
     # TODO: Discuss if this is a sensible defaut
@@ -1034,7 +1034,8 @@ class LanguageModelInfo(BaseModel):
                     encoder_name=EncoderName.O200K_BASE,
                     token_limits=LanguageModelTokenLimits(
                         # Limit of smallest underlying model; router can route to larger contexts
-                        token_limit_input=272_000, token_limit_output=32_768
+                        token_limit_input=272_000,
+                        token_limit_output=32_768,
                     ),
                 )
             case LanguageModelName.ANTHROPIC_CLAUDE_3_7_SONNET: