unique-toolkit 2026.34.0.dev10__tar.gz → 2026.34.0.dev12__tar.gz

{unique_toolkit-2026.34.0.dev10 → unique_toolkit-2026.34.0.dev12}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.3
 Name: unique-toolkit
-Version: 2026.34.0.dev10
+Version: 2026.34.0.dev12
 Summary: 
 Author: Cedric Klinkert, Martin Fadler, Sadique Sheik, Fabian Schläpfer, Pascal Hauri, Ejiroghene Oki-Peter, Rami Azouz, Ricardo Visini, Sanzio Monti
 Author-email: Cedric Klinkert <cedric.klinkert@unique.ch>, Martin Fadler <martin.fadler@unique.ch>, Sadique Sheik <sadique@unique.ch>, Fabian Schläpfer <fabian@unique.ch>, Pascal Hauri <pascal@unique.ch>, Ejiroghene Oki-Peter <oki-peter.ejiroghene@unique.ai>, Rami Azouz <rami.ext@unique.ch>, Ricardo Visini <ricardo.visini@unique.ch>, Sanzio Monti <sanzio.monti@unique.ai>
@@ -16,7 +16,7 @@ Requires-Dist: sseclient>=0.0.27,<0.0.28
 Requires-Dist: openai>=1.105.0,<3
 Requires-Dist: platformdirs>=4.0.0,<5
 Requires-Dist: pillow>=12.1.1
-Requires-Dist: unique-sdk>=2026.34.0.dev4,<2026.34.0rc0
+Requires-Dist: unique-sdk>=2026.34.0.dev5,<2026.34.0rc0
 Requires-Dist: jambo>=0.1.2,<0.2
 Requires-Dist: jinja2>=3.1.6,<4
 Requires-Dist: docxtpl>=0.20.1,<0.21

{unique_toolkit-2026.34.0.dev10 → unique_toolkit-2026.34.0.dev12}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "unique_toolkit"
-version = "2026.34.0.dev10"
+version = "2026.34.0.dev12"
 description = ""
 readme = "README.md"
 requires-python = ">=3.12"
@@ -28,7 +28,7 @@ dependencies = [
     "openai>=1.105.0,<3",
     "platformdirs>=4.0.0,<5",
     "pillow>=12.1.1",
-    "unique-sdk>=2026.34.0.dev4,<2026.34.0rc0",
+    "unique-sdk>=2026.34.0.dev5,<2026.34.0rc0",
     "jambo>=0.1.2,<0.2",
     "jinja2>=3.1.6,<4",
     "docxtpl>=0.20.1,<0.21",

{unique_toolkit-2026.34.0.dev10 → unique_toolkit-2026.34.0.dev12}/unique_toolkit/agentic/evaluation/hallucination/hallucination_evaluation.py

@@ -36,6 +36,8 @@ class HallucinationEvaluation(Evaluation):
         self.config = config
         self._company_id = event.company_id
         self._user_id = event.user_id
+        self._chat_id = event.payload.chat_id
+        self._assistant_id = event.payload.assistant_id
         self._reference_manager = reference_manager
         self._user_message = event.payload.user_message.text
         super().__init__(EvaluationMetricName.HALLUCINATION)
@@ -75,6 +77,8 @@ class HallucinationEvaluation(Evaluation):
                 output_text=loop_response.message.text,
             ),
             config=self.config,
+            chat_id=self._chat_id,
+            assistant_id=self._assistant_id,
         )
 
         score_to_label = self.config.score_to_label

{unique_toolkit-2026.34.0.dev10 → unique_toolkit-2026.34.0.dev12}/unique_toolkit/agentic/evaluation/hallucination/service.py

@@ -59,4 +59,6 @@ class HallucinationEvaluator:
             user_id=self.event.user_id,
             input=input,
             config=config,
+            chat_id=self.event.payload.chat_id,
+            assistant_id=self.event.payload.assistant_id,
         )

{unique_toolkit-2026.34.0.dev10 → unique_toolkit-2026.34.0.dev12}/unique_toolkit/agentic/evaluation/hallucination/utils.py

@@ -37,6 +37,8 @@ async def check_hallucination(
     user_id: str,
     input: EvaluationMetricInput,
     config: EvaluationMetricConfig,
+    chat_id: str | None = None,
+    assistant_id: str | None = None,
 ) -> EvaluationMetricResult:
     """
     Analyzes the level of hallucination in the generated output by comparing it with the provided input
@@ -77,7 +79,12 @@ async def check_hallucination(
     try:
         msgs = _get_msgs(input, config)
         result = await LanguageModelService.complete_async_util(
-            company_id=company_id, user_id=user_id, messages=msgs, model_name=model_name
+            company_id=company_id,
+            user_id=user_id,
+            messages=msgs,
+            model_name=model_name,
+            chat_id=chat_id,
+            assistant_id=assistant_id,
         )
     except Exception as e:
         error_message = "Error occurred during hallucination metric analysis"

{unique_toolkit-2026.34.0.dev10 → unique_toolkit-2026.34.0.dev12}/unique_toolkit/agentic/evaluation/tests/test_hallucination_external_context.py

@@ -36,6 +36,8 @@ def _make_event() -> SimpleNamespace:
         company_id="company_1",
         user_id="user_1",
         payload=SimpleNamespace(
+            chat_id="chat_1",
+            assistant_id="assistant_1",
             user_message=SimpleNamespace(text="what is ACME revenue?"),
         ),
     )
@@ -61,8 +63,18 @@ def captured_context(monkeypatch: pytest.MonkeyPatch) -> dict:
     """Patch check_hallucination to capture the context_texts it receives."""
     captured: dict = {}
 
-    async def _fake_check_hallucination(*, company_id, user_id, input, config):
+    async def _fake_check_hallucination(
+        *,
+        company_id,
+        user_id,
+        input,
+        config,
+        chat_id,
+        assistant_id,
+    ):
         captured["context_texts"] = list(input.context_texts or [])
+        captured["chat_id"] = chat_id
+        captured["assistant_id"] = assistant_id
         return EvaluationMetricResult(
             name=EvaluationMetricName.HALLUCINATION,
             value="LOW",
@@ -86,6 +98,8 @@ async def test_external_context_is_appended(captured_context: dict) -> None:
 
     # No chunks → context comes solely from the external (MCP) source.
     assert captured_context["context_texts"] == ["ACME revenue: 1M"]
+    assert captured_context["chat_id"] == "chat_1"
+    assert captured_context["assistant_id"] == "assistant_1"
     assert result.value == "LOW"
     assert result.is_positive is True
 

{unique_toolkit-2026.34.0.dev10 → unique_toolkit-2026.34.0.dev12}/unique_toolkit/agentic/evaluation/tests/test_hallucination_utils.py

@@ -1,6 +1,7 @@
 """Tests for hallucination evaluation utils."""
 
 from typing import List, Optional
+from unittest.mock import AsyncMock
 
 import pytest
 
@@ -14,6 +15,7 @@ from unique_toolkit.agentic.evaluation.hallucination.utils import (
     _from_order_source_selection_mode,
     _from_original_response_source_selection_mode,
     _get_msgs,
+    check_hallucination,
     context_text_from_stream_response,
 )
 from unique_toolkit.agentic.evaluation.schemas import EvaluationMetricInput
@@ -25,8 +27,11 @@ from unique_toolkit.language_model.infos import (
     ModelCapabilities,
 )
 from unique_toolkit.language_model.schemas import (
+    LanguageModelAssistantMessage,
+    LanguageModelCompletionChoice,
     LanguageModelMessageRole,
     LanguageModelMessages,
+    LanguageModelResponse,
     LanguageModelStreamResponse,
 )
 
@@ -101,6 +106,49 @@ def evaluation_input() -> EvaluationMetricInput:
     )
 
 
+@pytest.mark.ai
+@pytest.mark.asyncio
+async def test_check_hallucination__forwards_attribution_ids(
+    monkeypatch: pytest.MonkeyPatch,
+    hallucination_config: HallucinationConfig,
+    evaluation_input: EvaluationMetricInput,
+) -> None:
+    """Purpose: Verify hallucination completion receives chat attribution IDs.
+    Why this matters: Secondary evaluator usage must link to the originating chat.
+    Setup summary: Mock the completion utility, run the check, and inspect its call.
+    """
+    complete_async_util = AsyncMock(
+        return_value=LanguageModelResponse(
+            choices=[
+                LanguageModelCompletionChoice(
+                    index=0,
+                    message=LanguageModelAssistantMessage(
+                        content='{"value": "low", "reason": "grounded"}'
+                    ),
+                    finish_reason="stop",
+                )
+            ]
+        )
+    )
+    monkeypatch.setattr(
+        "unique_toolkit.agentic.evaluation.hallucination.utils."
+        "LanguageModelService.complete_async_util",
+        complete_async_util,
+    )
+
+    await check_hallucination(
+        company_id="company_1",
+        user_id="user_1",
+        input=evaluation_input.model_copy(update={"history_messages": []}),
+        config=hallucination_config,
+        chat_id="chat_1",
+        assistant_id="assistant_1",
+    )
+
+    assert complete_async_util.await_args.kwargs["chat_id"] == "chat_1"
+    assert complete_async_util.await_args.kwargs["assistant_id"] == "assistant_1"
+
+
 @pytest.mark.ai
 def test_default_source_selection_mode__selects_chunks__by_source_id_match(
     sample_references: List[ContentReference],

{unique_toolkit-2026.34.0.dev10 → unique_toolkit-2026.34.0.dev12}/unique_toolkit/agentic/history_manager/history_manager.py

@@ -348,8 +348,6 @@ class HistoryManager:
             int(m)
             for m in re.findall(r"\[source(\d+)\]", assistant_text, re.IGNORECASE)
         }
-        if not cited:
-            return records
 
         return [
             record.model_copy(
@@ -374,8 +372,9 @@ def _strip_uncited_sources_from_content(content: str, cited: set[int]) -> str:
 
     The content is expected to be a JSON array of dicts containing at least
     a "source_number" key.  Items whose "source_number" is not in *cited*
-    are removed.  If the content is not valid JSON or not in the expected
-    format it is returned unchanged.
+    are removed. If no sources are cited, the response is replaced with a short
+    message. If the content is not valid JSON or not in the expected format it
+    is returned unchanged.
     """
     try:
         data = json.loads(content)
@@ -390,5 +389,10 @@ def _strip_uncited_sources_from_content(content: str, cited: set[int]) -> str:
     ):
         return content
 
+    if not cited:
+        return (
+            "The search did not produce any relevant information that could be cited."
+        )
+
     filtered = [item for item in data if item.get("source_number") in cited]
     return serialize_tool_content_json(filtered)

{unique_toolkit-2026.34.0.dev10 → unique_toolkit-2026.34.0.dev12}/unique_toolkit/language_model/functions.py

@@ -41,6 +41,18 @@ from .constants import (
 logger = logging.getLogger(f"toolkit.language_model.{__name__}")
 
 
+def _completion_headers(
+    chat_id: str | None,
+    assistant_id: str | None,
+) -> dict[str, str] | None:
+    headers: dict[str, str] = {}
+    if chat_id:
+        headers["x-chat-id"] = chat_id
+    if assistant_id:
+        headers["x-assistant-id"] = assistant_id
+    return headers or None
+
+
 def complete(
     company_id: str,
     messages: LanguageModelMessages | list[ChatCompletionMessageParam],
@@ -52,6 +64,8 @@ def complete(
     structured_output_model: type[BaseModel] | dict[str, Any] | None = None,
     structured_output_enforce_schema: bool = False,
     user_id: str | None = None,
+    chat_id: str | None = None,
+    assistant_id: str | None = None,
 ) -> LanguageModelResponse:
     """Call the completion endpoint synchronously without streaming the response.
 
@@ -87,14 +101,26 @@ def complete(
         )
 
     try:
-        response = unique_sdk.ChatCompletion.create(
-            company_id=company_id,
-            user_id=user_id,
-            model=model,  # pyright: ignore[reportArgumentType]
-            messages=messages_dict,  # pyright: ignore[reportArgumentType]
-            timeout=timeout,
-            options=options,
-        )
+        headers = _completion_headers(chat_id, assistant_id)
+        if headers is None:
+            response = unique_sdk.ChatCompletion.create(
+                company_id=company_id,
+                user_id=user_id,
+                model=model,  # pyright: ignore[reportArgumentType]
+                messages=messages_dict,  # pyright: ignore[reportArgumentType]
+                timeout=timeout,
+                options=options,
+            )
+        else:
+            response = unique_sdk.ChatCompletion.create(
+                company_id=company_id,
+                user_id=user_id,
+                headers=headers,
+                model=model,  # pyright: ignore[reportArgumentType]
+                messages=messages_dict,  # pyright: ignore[reportArgumentType]
+                timeout=timeout,
+                options=options,
+            )
         return LanguageModelResponse(**response)
     except Exception as e:
         logger.error(f"Error completing: {e}")
@@ -112,6 +138,8 @@ async def complete_async(
     other_options: dict[str, Any] | None = None,
     structured_output_model: type[BaseModel] | dict[str, Any] | None = None,
     structured_output_enforce_schema: bool = False,
+    chat_id: str | None = None,
+    assistant_id: str | None = None,
 ) -> LanguageModelResponse:
     """Call the completion endpoint asynchronously without streaming the response.
 
@@ -156,14 +184,26 @@ async def complete_async(
     )
 
     try:
-        response = await unique_sdk.ChatCompletion.create_async(
-            company_id=company_id,
-            user_id=user_id,
-            model=model,  # pyright: ignore[reportArgumentType]
-            messages=messages_dict,  # pyright: ignore[reportArgumentType]
-            timeout=timeout,
-            options=options,
-        )
+        headers = _completion_headers(chat_id, assistant_id)
+        if headers is None:
+            response = await unique_sdk.ChatCompletion.create_async(
+                company_id=company_id,
+                user_id=user_id,
+                model=model,  # pyright: ignore[reportArgumentType]
+                messages=messages_dict,  # pyright: ignore[reportArgumentType]
+                timeout=timeout,
+                options=options,
+            )
+        else:
+            response = await unique_sdk.ChatCompletion.create_async(
+                company_id=company_id,
+                user_id=user_id,
+                headers=headers,
+                model=model,  # pyright: ignore[reportArgumentType]
+                messages=messages_dict,  # pyright: ignore[reportArgumentType]
+                timeout=timeout,
+                options=options,
+            )
         return LanguageModelResponse(**response)
     except Exception as e:
         logger.exception(f"Error completing: {e}")
@@ -434,6 +474,8 @@ def complete_with_references(
     tools: list[LanguageModelTool | LanguageModelToolDescription] | None = None,
     start_text: str | None = None,
     other_options: dict[str, Any] | None = None,
+    chat_id: str | None = None,
+    assistant_id: str | None = None,
 ) -> LanguageModelStreamResponse:
     # Use toolkit language model functions for chat completion
 
@@ -446,6 +488,8 @@ def complete_with_references(
         timeout=timeout,
         tools=tools,
         other_options=other_options,
+        chat_id=chat_id,
+        assistant_id=assistant_id,
     )
 
     return _create_language_model_stream_response_with_references(
@@ -467,6 +511,8 @@ async def complete_with_references_async(
     tools: list[LanguageModelTool | LanguageModelToolDescription] | None = None,
     start_text: str | None = None,
     other_options: dict[str, Any] | None = None,
+    chat_id: str | None = None,
+    assistant_id: str | None = None,
 ) -> LanguageModelStreamResponse:
     # Use toolkit language model functions for chat completion
     response = await complete_async(
@@ -478,6 +524,8 @@ async def complete_with_references_async(
         timeout=timeout,
         tools=tools,
         other_options=other_options,
+        chat_id=chat_id,
+        assistant_id=assistant_id,
     )
 
     return _create_language_model_stream_response_with_references(

{unique_toolkit-2026.34.0.dev10 → unique_toolkit-2026.34.0.dev12}/unique_toolkit/language_model/service.py

@@ -85,7 +85,11 @@ class LanguageModelService:
         """
         Initialize the LanguageModelService with an event.
         """
-        return cls(company_id=event.company_id, user_id=event.user_id)
+        service = cls(company_id=event.company_id, user_id=event.user_id)
+        if isinstance(event, AssistantWebhookEvent):
+            service._chat_id = event.payload.chat_id
+            service._assistant_id = event.payload.assistant_id
+        return service
 
     @classmethod
     def from_settings(cls, settings: UniqueSettings | str | None = None, **kwargs: Any):
@@ -245,6 +249,8 @@ class LanguageModelService:
             other_options=other_options,
             structured_output_model=structured_output_model,
             structured_output_enforce_schema=structured_output_enforce_schema,
+            chat_id=self._chat_id,
+            assistant_id=self._assistant_id,
         )
 
     async def complete_async(
@@ -273,6 +279,8 @@ class LanguageModelService:
             other_options=other_options,
             structured_output_model=structured_output_model,
             structured_output_enforce_schema=structured_output_enforce_schema,
+            chat_id=self._chat_id,
+            assistant_id=self._assistant_id,
         )
 
     @classmethod
@@ -289,6 +297,8 @@ class LanguageModelService:
         structured_output_model: Optional[Type[BaseModel] | dict[str, Any]] = None,
         structured_output_enforce_schema: bool = False,
         other_options: Optional[dict[str, Any]] = None,
+        chat_id: str | None = None,
+        assistant_id: str | None = None,
     ) -> LanguageModelResponse:
         """
         Calls the completion endpoint asynchronously without streaming the response.
@@ -305,6 +315,8 @@ class LanguageModelService:
             other_options=other_options,
             structured_output_model=structured_output_model,
             structured_output_enforce_schema=structured_output_enforce_schema,
+            chat_id=chat_id,
+            assistant_id=assistant_id,
         )
 
     def complete_with_references(
@@ -330,6 +342,8 @@ class LanguageModelService:
             other_options=other_options,
             tools=tools,
             start_text=start_text,
+            chat_id=self._chat_id,
+            assistant_id=self._assistant_id,
         )
 
     async def complete_with_references_async(
@@ -355,4 +369,6 @@ class LanguageModelService:
             other_options=other_options,
             tools=tools,
             start_text=start_text,
+            chat_id=self._chat_id,
+            assistant_id=self._assistant_id,
         )