unique_toolkit 1.46.1__tar.gz → 1.46.3__tar.gz

{unique_toolkit-1.46.1 → unique_toolkit-1.46.3}/CHANGELOG.md

@@ -5,6 +5,15 @@ 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).
 
+
+## [1.46.3] - 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)
+
 ## [1.46.1] - 2026-02-06
 - Add model-agnostic token counting via `LanguageModelInfo.get_encoder()`
 - Add bundled Qwen/DeepSeek tokenizers for accurate token counting

{unique_toolkit-1.46.1 → unique_toolkit-1.46.3}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: unique_toolkit
-Version: 1.46.1
+Version: 1.46.3
 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 +125,15 @@ 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).
 
+
+## [1.46.3] - 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)
+
 ## [1.46.1] - 2026-02-06
 - Add model-agnostic token counting via `LanguageModelInfo.get_encoder()`
 - Add bundled Qwen/DeepSeek tokenizers for accurate token counting

{unique_toolkit-1.46.1 → unique_toolkit-1.46.3}/pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "unique_toolkit"
-version = "1.46.1"
+version = "1.46.3"
 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-1.46.1 → unique_toolkit-1.46.3}/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-1.46.1 → unique_toolkit-1.46.3}/unique_toolkit/language_model/infos.py

@@ -49,6 +49,7 @@ class LanguageModelName(StrEnum):
     AZURE_GPT_41_NANO_2025_0414 = "AZURE_GPT_41_NANO_2025_0414"
     AZURE_o3_2025_0416 = "AZURE_o3_2025_0416"
     AZURE_o4_MINI_2025_0416 = "AZURE_o4_MINI_2025_0416"
+    AZURE_MODEL_ROUTER_2025_1118 = "AZURE_MODEL_ROUTER_2025_1118"
     ANTHROPIC_CLAUDE_3_7_SONNET = "litellm:anthropic-claude-3-7-sonnet"
     ANTHROPIC_CLAUDE_3_7_SONNET_THINKING = (
         "litellm:anthropic-claude-3-7-sonnet-thinking"
@@ -59,6 +60,7 @@ class LanguageModelName(StrEnum):
     ANTHROPIC_CLAUDE_OPUS_4 = "litellm:anthropic-claude-opus-4"
     ANTHROPIC_CLAUDE_OPUS_4_1 = "litellm:anthropic-claude-opus-4-1"
     ANTHROPIC_CLAUDE_OPUS_4_5 = "litellm:anthropic-claude-opus-4-5"
+    ANTHROPIC_CLAUDE_OPUS_4_6 = "litellm:anthropic-claude-opus-4-6"
     GEMINI_2_0_FLASH = "litellm:gemini-2-0-flash"
     GEMINI_2_5_FLASH = "litellm:gemini-2-5-flash"
     GEMINI_2_5_FLASH_LITE = "litellm:gemini-2-5-flash-lite"
@@ -142,6 +144,7 @@ def get_encoder_name(model_name: LanguageModelName) -> EncoderName:
             | LMN.AZURE_GPT_51_CODEX_MINI_2025_1113
             | LMN.AZURE_GPT_52_2025_1211
             | LMN.AZURE_GPT_52_CHAT_2025_1211
+            | LMN.AZURE_MODEL_ROUTER_2025_1118
             | LMN.LITELLM_OPENAI_GPT_5
             | LMN.LITELLM_OPENAI_GPT_5_MINI
             | LMN.LITELLM_OPENAI_GPT_5_NANO
@@ -1015,6 +1018,26 @@ class LanguageModelInfo(BaseModel):
                     info_cutoff_at=date(2024, 5, 31),
                     published_at=date(2025, 4, 14),
                 )
+            case LanguageModelName.AZURE_MODEL_ROUTER_2025_1118:
+                return cls(
+                    name=model_name,
+                    capabilities=[
+                        ModelCapabilities.CHAT_COMPLETIONS_API,
+                        ModelCapabilities.FUNCTION_CALLING,
+                        ModelCapabilities.RESPONSES_API,
+                        ModelCapabilities.STREAMING,
+                        ModelCapabilities.STRUCTURED_OUTPUT,
+                        ModelCapabilities.VISION,
+                    ],
+                    provider=LanguageModelProvider.AZURE,
+                    version="2025-11-18",
+                    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,
+                    ),
+                )
             case LanguageModelName.ANTHROPIC_CLAUDE_3_7_SONNET:
                 return cls(
                     name=model_name,
@@ -1174,6 +1197,26 @@ class LanguageModelInfo(BaseModel):
                     info_cutoff_at=date(2025, 8, 1),
                     published_at=date(2025, 11, 13),
                 )
+            case LanguageModelName.ANTHROPIC_CLAUDE_OPUS_4_6:
+                return cls(
+                    name=model_name,
+                    capabilities=[
+                        ModelCapabilities.FUNCTION_CALLING,
+                        ModelCapabilities.STREAMING,
+                        ModelCapabilities.VISION,
+                        ModelCapabilities.REASONING,
+                    ],
+                    provider=LanguageModelProvider.LITELLM,
+                    version="claude-opus-4-6",
+                    encoder_name=EncoderName.O200K_BASE,  # TODO: Update encoder with litellm
+                    token_limits=LanguageModelTokenLimits(
+                        # Input limit is 200_000, we leave 20_000 tokens as buffer due to tokenizer mismatch
+                        token_limit_input=180_000,
+                        token_limit_output=128_000,
+                    ),
+                    info_cutoff_at=date(2025, 8, 1),
+                    published_at=date(2026, 2, 5),
+                )
             case LanguageModelName.GEMINI_2_0_FLASH:
                 return cls(
                     name=model_name,