unique_toolkit 2026.44.0.dev2__tar.gz → 2026.44.0.dev4__tar.gz

{unique_toolkit-2026.44.0.dev2 → unique_toolkit-2026.44.0.dev4}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.3
 Name: unique-toolkit
-Version: 2026.44.0.dev2
+Version: 2026.44.0.dev4
 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>=2026.44.0.dev1,<2026.44.0rc0
+Requires-Dist: unique-sdk>=2026.44.0.dev7,<2026.44.0rc0
 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-2026.44.0.dev2 → unique_toolkit-2026.44.0.dev4}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "unique_toolkit"
-version = "2026.44.0.dev2"
+version = "2026.44.0.dev4"
 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>=2026.44.0.dev1,<2026.44.0rc0",
+    "unique-sdk>=2026.44.0.dev7,<2026.44.0rc0",
     "jambo>=0.1.2,<0.2",
     "jinja2>=3.1.6,<4",
     "docxtpl>=0.20.1,<0.21",

{unique_toolkit-2026.44.0.dev2 → unique_toolkit-2026.44.0.dev4}/unique_toolkit/content/__init__.py

@@ -16,6 +16,9 @@ from .schemas import (
 from .schemas import (
     ContentRerankerConfig as ContentRerankerConfig,
 )
+from .schemas import (
+    ContentSearchMergeStrategy as ContentSearchMergeStrategy,
+)
 from .schemas import (
     ContentSearchResult as ContentSearchResult,
 )
@@ -47,6 +50,7 @@ __all__ = [
     "ContentMetadata",
     "ContentReference",
     "ContentRerankerConfig",
+    "ContentSearchMergeStrategy",
     "ContentSearchResult",
     "ContentSearchType",
     "ContentService",

{unique_toolkit-2026.44.0.dev2 → unique_toolkit-2026.44.0.dev4}/unique_toolkit/content/functions.py

@@ -17,6 +17,7 @@ from unique_toolkit.content.schemas import (
     ContentChunk,
     ContentInfo,
     ContentRerankerConfig,
+    ContentSearchMergeStrategy,
     ContentSearchType,
     DeleteContentResponse,
     FolderInfo,
@@ -45,6 +46,8 @@ def search_content_chunks(
     metadata_filter: dict[str, Any] | None = None,
     content_ids: list[str] | None = None,
     score_threshold: float | None = None,
+    merge_strategy: ContentSearchMergeStrategy | None = None,
+    include_scores: bool | None = None,
 ) -> list[ContentChunk]:
     """
     Performs a synchronous search for content chunks in the knowledge base.
@@ -60,6 +63,8 @@ def search_content_chunks(
         metadata_filter (dict | None): UniqueQL metadata filter. If unspecified/None, it tries to use the metadata filter from the event. Defaults to None.
         content_ids (list[str] | None): The content IDs to search. Defaults to None.
         score_threshold (float | None): The minimum score threshold for results. Defaults to 0.
+        merge_strategy (ContentSearchMergeStrategy | None): How a COMBINED search merges vector and full-text results. ``ZIP`` interleaves the two lists and keeps each result's original score. ``RRF`` and ``DBSF`` replace those scores with a fused score. Ignored by other search types. Defaults to the backend configuration.
+        include_scores (bool | None): When true, each ``ContentChunk.score`` is set. Vector search returns a similarity score, full-text search a BM25 score, and a COMBINED search a fused score for ``RRF`` and ``DBSF``. Scores are only comparable within one search type. Defaults to None.
     Returns:
         list[ContentChunk]: The search results.
     """
@@ -86,6 +91,8 @@ def search_content_chunks(
             metaDataFilter=metadata_filter,
             contentIds=content_ids,
             scoreThreshold=score_threshold,
+            mergeStrategy=merge_strategy.value if merge_strategy else None,
+            includeScores=include_scores,
         )
         return map_to_content_chunks(searches)
     except Exception as e:
@@ -107,6 +114,8 @@ async def search_content_chunks_async(
     metadata_filter: dict[str, Any] | None = None,
     content_ids: list[str] | None = None,
     score_threshold: float | None = None,
+    merge_strategy: ContentSearchMergeStrategy | None = None,
+    include_scores: bool | None = None,
 ):
     """
     Performs an asynchronous search for content chunks in the knowledge base.
@@ -136,6 +145,8 @@ async def search_content_chunks_async(
             metaDataFilter=metadata_filter,
             contentIds=content_ids,
             scoreThreshold=score_threshold,
+            mergeStrategy=merge_strategy.value if merge_strategy else None,
+            includeScores=include_scores,
         )
         return map_to_content_chunks(searches)
     except Exception as e:

{unique_toolkit-2026.44.0.dev2 → unique_toolkit-2026.44.0.dev4}/unique_toolkit/content/schemas.py

@@ -70,6 +70,11 @@ class ContentChunk(BaseModel):
         description="The end page of the chunk. For document chunks this is the end page of the document.",
     )
 
+    score: float | None = Field(
+        default=None,
+        description="The relevance score of the chunk. Only set when the search requested scores. Scores are comparable within a single search type only.",
+    )
+
     object: str | None = None
     metadata: ContentMetadata | None = None
     internally_stored_at: datetime | None = None
@@ -294,6 +299,18 @@ class ContentSearchType(StrEnum):
     VECTOR = "VECTOR"
 
 
+class ContentSearchMergeStrategy(StrEnum):
+    """How a ``COMBINED`` search merges vector and full-text results.
+
+    Ignored by single-source search types. When omitted, the backend applies its
+    configured default.
+    """
+
+    ZIP = "ZIP"
+    RRF = "RRF"
+    DBSF = "DBSF"
+
+
 class ContentSearchResult(BaseModel):
     """Schema corresponding to unique_sdk.SearchResult"""
 

{unique_toolkit-2026.44.0.dev2 → unique_toolkit-2026.44.0.dev4}/unique_toolkit/content/service.py

@@ -30,6 +30,7 @@ from unique_toolkit.content.schemas import (
     Content,
     ContentChunk,
     ContentRerankerConfig,
+    ContentSearchMergeStrategy,
     ContentSearchType,
 )
 
@@ -357,6 +358,8 @@ class ContentService:
         metadata_filter: dict[str, Any] | None = None,
         content_ids: list[str] | None = None,
         score_threshold: float | None = None,
+        merge_strategy: ContentSearchMergeStrategy | None = None,
+        include_scores: bool | None = None,
     ) -> list[ContentChunk]:
         """
         Performs a synchronous search for content chunks in the knowledge base.
@@ -373,6 +376,8 @@ class ContentService:
             metadata_filter (dict | None, optional): UniqueQL metadata filter. If unspecified/None, it tries to use the metadata filter from the event. Defaults to None.
             content_ids (list[str] | None, optional): The content IDs to search within. Defaults to None.
             score_threshold (float | None, optional): Sets the minimum similarity score for search results to be considered. Defaults to 0.
+            merge_strategy (ContentSearchMergeStrategy | None, optional): How a COMBINED search merges vector and full-text results. ``ZIP`` interleaves the two lists and keeps each result's original score. ``RRF`` and ``DBSF`` replace those scores with a fused score. Ignored by other search types. Defaults to the backend configuration.
+            include_scores (bool | None, optional): When true, each ``ContentChunk.score`` is set. Vector search returns a similarity score, full-text search a BM25 score, and a COMBINED search a fused score for ``RRF`` and ``DBSF``. Scores are only comparable within one search type. Defaults to None.
 
         Returns:
             list[ContentChunk]: The search results.
@@ -404,6 +409,8 @@ class ContentService:
                 metadata_filter=metadata_filter,
                 content_ids=content_ids,
                 score_threshold=score_threshold,
+                merge_strategy=merge_strategy,
+                include_scores=include_scores,
             )
             return searches
         except Exception as e:
@@ -424,6 +431,8 @@ class ContentService:
         metadata_filter: dict[str, Any] | None = None,
         content_ids: list[str] | None = None,
         score_threshold: float | None = None,
+        merge_strategy: ContentSearchMergeStrategy | None = None,
+        include_scores: bool | None = None,
     ):
         """
         Performs an asynchronous search for content chunks in the knowledge base.
@@ -440,6 +449,8 @@ class ContentService:
             metadata_filter (dict | None, optional): UniqueQL metadata filter. If unspecified/None, it tries to use the metadata filter from the event. Defaults to None.
             content_ids (list[str] | None, optional): The content IDs to search within. Defaults to None.
             score_threshold (float | None, optional): Sets the minimum similarity score for search results to be considered. Defaults to 0.
+            merge_strategy (ContentSearchMergeStrategy | None, optional): How a COMBINED search merges vector and full-text results. ``ZIP`` interleaves the two lists and keeps each result's original score. ``RRF`` and ``DBSF`` replace those scores with a fused score. Ignored by other search types. Defaults to the backend configuration.
+            include_scores (bool | None, optional): When true, each ``ContentChunk.score`` is set. Vector search returns a similarity score, full-text search a BM25 score, and a COMBINED search a fused score for ``RRF`` and ``DBSF``. Scores are only comparable within one search type. Defaults to None.
 
         Returns:
             list[ContentChunk]: The search results.
@@ -470,6 +481,8 @@ class ContentService:
                 metadata_filter=metadata_filter,
                 content_ids=content_ids,
                 score_threshold=score_threshold,
+                merge_strategy=merge_strategy,
+                include_scores=include_scores,
             )
             return searches
         except Exception as e:

{unique_toolkit-2026.44.0.dev2 → unique_toolkit-2026.44.0.dev4}/unique_toolkit/language_model/infos.py

@@ -176,6 +176,7 @@ class LanguageModelName(StrEnum):
     VERTEX_CLAUDE_OPUS_5_5 = "litellm:vertex-claude-opus-5-5"
     VERTEX_CLAUDE_FABLE_5 = "litellm:vertex-claude-fable-5"
     VERTEX_GEMINI_3_8_FLASH = "litellm:vertex-gemini-3-8-flash"
+    VERTEX_GEMINI_4_ARGON = "litellm:vertex-gemini-4-argon"
 
 
 class EncoderName(StrEnum):
@@ -2487,11 +2488,10 @@ class LanguageModelInfo(BaseModel):
                     default_options={"reasoning_effort": "low"},
                     supported_reasoning_efforts=["low", "high"],
                 )
-            # Day-0 placeholder so the model can be selected before general
-            # availability. Limits, cutoff, thinking levels, and the API model
-            # id are provisional. Update this entry from the official model
-            # card once Argon is published.
-            case LanguageModelName.GEMINI_4_ARGON:
+            case (
+                LanguageModelName.GEMINI_4_ARGON
+                | LanguageModelName.VERTEX_GEMINI_4_ARGON
+            ):
                 return cls(
                     name=model_name,
                     capabilities=[
@@ -2506,21 +2506,16 @@ class LanguageModelInfo(BaseModel):
                     version="gemini-4-argon",
                     encoder_name=EncoderName.O200K_BASE,  # TODO: Update encoder with litellm
                     token_limits=LanguageModelTokenLimits(
-                        # GraphWalks evaluates contexts from 256k through 1M tokens.
-                        # https://deepmind.google/models/gemini/
+                        # Unified 1,048,576 input/output context window; 65,536 single-request output cap.
+                        # https://ai.google.dev/gemini-api/docs/models/gemini-4-argon
+                        # https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/gemini/4-argon
                         token_limit_input=1_048_576,
                         token_limit_output=65_536,
                     ),
-                    # Google has not published a knowledge cutoff for Argon.
-                    info_cutoff_at=None,
-                    # https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon/
-                    published_at=date(2026, 9, 30),
-                    # thinkingLevel is recommended for Gemini 3 and later. Argon is
-                    # not yet on the public thinking-level table; evals use the
-                    # highest thinking setting. Mirror Gemini 3.8 Flash (low,
-                    # medium, high; medium default) until Argon docs land.
-                    # https://ai.google.dev/gemini-api/docs/thinking
-                    # https://deepmind.google/models/evals-methodology/gemini-4-argon
+                    info_cutoff_at=date(2026, 3, day=1),
+                    published_at=date(2026, 10, 13),
+                    # Supports thinking_level low, medium (default), and high.
+                    # https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/guides/gemini-4-argon
                     default_options={"reasoning_effort": "medium"},
                     supported_reasoning_efforts=["low", "medium", "high"],
                 )

{unique_toolkit-2026.44.0.dev2 → unique_toolkit-2026.44.0.dev4}/unique_toolkit/services/chat_service.py

@@ -97,6 +97,7 @@ from unique_toolkit.content.schemas import (
     ContentChunk,
     ContentReference,
     ContentRerankerConfig,
+    ContentSearchMergeStrategy,
     ContentSearchType,
 )
 from unique_toolkit.elicitation.service import ElicitationService
@@ -2119,8 +2120,15 @@ class ChatService(ChatServiceDeprecated):
         metadata_filter: dict[str, Any] | None = None,
         content_ids: list[str] | None = None,
         score_threshold: float | None = None,
+        merge_strategy: ContentSearchMergeStrategy | None = None,
+        include_scores: bool | None = None,
     ) -> list[ContentChunk]:
-        """Search content chunks scoped to this chat session (chat_only=True)."""
+        """Search content chunks scoped to this chat session (chat_only=True).
+
+        Args:
+            merge_strategy: How a COMBINED search merges vector and full-text results. ``ZIP`` interleaves the two lists and keeps each result's original score. ``RRF`` and ``DBSF`` replace those scores with a fused score. Ignored by other search types. Defaults to the backend configuration.
+            include_scores: When true, each ``ContentChunk.score`` is set. Vector search returns a similarity score, full-text search a BM25 score, and a COMBINED search a fused score for ``RRF`` and ``DBSF``. Scores are only comparable within one search type.
+        """
         if scope_ids:
             warnings.warn(
                 "Passing scope_ids to ChatService.search_content_chunks_async is "
@@ -2147,6 +2155,8 @@ class ChatService(ChatServiceDeprecated):
             metadata_filter=metadata_filter,
             content_ids=content_ids,
             score_threshold=score_threshold,
+            merge_strategy=merge_strategy,
+            include_scores=include_scores,
         )
 
     async def search_contents_async(

{unique_toolkit-2026.44.0.dev2 → unique_toolkit-2026.44.0.dev4}/unique_toolkit/services/knowledge_base.py

@@ -46,6 +46,7 @@ from unique_toolkit.content.schemas import (
     ContentChunk,
     ContentInfo,
     ContentRerankerConfig,
+    ContentSearchMergeStrategy,
     ContentSearchType,
     DeleteContentResponse,
     FolderInfo,
@@ -159,6 +160,8 @@ class KnowledgeBaseService:
         score_threshold: float = _DEFAULT_SCORE_THRESHOLD,
         search_language: str = DEFAULT_SEARCH_LANGUAGE,
         reranker_config: ContentRerankerConfig | None = None,
+        merge_strategy: ContentSearchMergeStrategy | None = None,
+        include_scores: bool | None = None,
     ) -> list[ContentChunk]: ...
 
     @overload
@@ -173,6 +176,8 @@ class KnowledgeBaseService:
         score_threshold: float = _DEFAULT_SCORE_THRESHOLD,
         search_language: str = DEFAULT_SEARCH_LANGUAGE,
         reranker_config: ContentRerankerConfig | None = None,
+        merge_strategy: ContentSearchMergeStrategy | None = None,
+        include_scores: bool | None = None,
     ) -> list[ContentChunk]: ...
 
     def search_content_chunks(
@@ -187,6 +192,8 @@ class KnowledgeBaseService:
         metadata_filter: dict[str, Any] | None = None,
         content_ids: list[str] | None = None,
         score_threshold: float | None = None,
+        merge_strategy: ContentSearchMergeStrategy | None = None,
+        include_scores: bool | None = None,
     ) -> list[ContentChunk]:
         """
         Performs a synchronous search for content chunks in the knowledge base.
@@ -202,6 +209,8 @@ class KnowledgeBaseService:
             metadata_filter (dict | None, optional): UniqueQL metadata filter. If unspecified/None, it tries to use the metadata filter from the event. Defaults to None.
             content_ids (list[str] | None, optional): The content IDs to search within. Defaults to None.
             score_threshold (float | None, optional): Sets the minimum similarity score for search results to be considered. Defaults to 0.
+            merge_strategy (ContentSearchMergeStrategy | None, optional): How a COMBINED search merges vector and full-text results. ``ZIP`` interleaves the two lists and keeps each result's original score. ``RRF`` and ``DBSF`` replace those scores with a fused score. Ignored by other search types. Defaults to the backend configuration.
+            include_scores (bool | None, optional): When true, each ``ContentChunk.score`` is set. Vector search returns a similarity score, full-text search a BM25 score, and a COMBINED search a fused score for ``RRF`` and ``DBSF``. Scores are only comparable within one search type. Defaults to None.
 
         Returns:
             list[ContentChunk]: The search results.
@@ -240,6 +249,8 @@ class KnowledgeBaseService:
                 metadata_filter=metadata_filter,
                 content_ids=content_ids,
                 score_threshold=score_threshold,
+                merge_strategy=merge_strategy,
+                include_scores=include_scores,
             )
             return searches
         except Exception as e:
@@ -257,6 +268,8 @@ class KnowledgeBaseService:
         score_threshold: float = _DEFAULT_SCORE_THRESHOLD,
         search_language: str = DEFAULT_SEARCH_LANGUAGE,
         reranker_config: ContentRerankerConfig | None = None,
+        merge_strategy: ContentSearchMergeStrategy | None = None,
+        include_scores: bool | None = None,
     ) -> list[ContentChunk]: ...
 
     @overload
@@ -271,6 +284,8 @@ class KnowledgeBaseService:
         score_threshold: float = _DEFAULT_SCORE_THRESHOLD,
         search_language: str = DEFAULT_SEARCH_LANGUAGE,
         reranker_config: ContentRerankerConfig | None = None,
+        merge_strategy: ContentSearchMergeStrategy | None = None,
+        include_scores: bool | None = None,
     ) -> list[ContentChunk]: ...
 
     async def search_content_chunks_async(
@@ -285,6 +300,8 @@ class KnowledgeBaseService:
         metadata_filter: dict[str, Any] | None = None,
         content_ids: list[str] | None = None,
         score_threshold: float | None = None,
+        merge_strategy: ContentSearchMergeStrategy | None = None,
+        include_scores: bool | None = None,
     ):
         """
         Performs an asynchronous search for content chunks in the knowledge base.
@@ -300,6 +317,8 @@ class KnowledgeBaseService:
             metadata_filter (dict | None, optional): UniqueQL metadata filter. If unspecified/None, it tries to use the metadata filter from the event. Defaults to None.
             content_ids (list[str] | None, optional): The content IDs to search within. Defaults to None.
             score_threshold (float | None, optional): Sets the minimum similarity score for search results to be considered. Defaults to 0.
+            merge_strategy (ContentSearchMergeStrategy | None, optional): How a COMBINED search merges vector and full-text results. ``ZIP`` interleaves the two lists and keeps each result's original score. ``RRF`` and ``DBSF`` replace those scores with a fused score. Ignored by other search types. Defaults to the backend configuration.
+            include_scores (bool | None, optional): When true, each ``ContentChunk.score`` is set. Vector search returns a similarity score, full-text search a BM25 score, and a COMBINED search a fused score for ``RRF`` and ``DBSF``. Scores are only comparable within one search type. Defaults to None.
 
         Returns:
             list[ContentChunk]: The search results.
@@ -337,6 +356,8 @@ class KnowledgeBaseService:
                 metadata_filter=metadata_filter,
                 content_ids=content_ids,
                 score_threshold=score_threshold,
+                merge_strategy=merge_strategy,
+                include_scores=include_scores,
             )
             return searches
         except Exception as e: