unique_web_search 2026.44.0.dev4__tar.gz → 2026.44.0.dev6__tar.gz

{unique_web_search-2026.44.0.dev4 → unique_web_search-2026.44.0.dev6}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.3
 Name: unique-web-search
-Version: 2026.44.0.dev4
+Version: 2026.44.0.dev6
 Summary: 
 Author: Andreas Hauri, Gustav Hartz, Rami Azouz
 Author-email: Andreas Hauri <andreas@unique.ch>, Gustav Hartz <gustav.hartz.ext@unique.ch>, Rami Azouz <rami.ext@unique.ch>
@@ -16,7 +16,7 @@ Requires-Dist: tavily-python>=0.7.11,<1
 Requires-Dist: unidecode>=1.4.0,<2
 Requires-Dist: azure-ai-projects>=2.3,<3
 Requires-Dist: azure-identity>=1.25.0,<2
-Requires-Dist: unique-toolkit[monitoring]>=2026.44.0.dev9,<2026.44.0rc0
+Requires-Dist: unique-toolkit[monitoring]>=2026.44.0.dev10,<2026.44.0rc0
 Requires-Dist: azure-core>=1.36.0,<2
 Requires-Dist: google-cloud-aiplatform>=1.128.0,<2
 Requires-Dist: google-auth>=2.43.0,<3
@@ -25,8 +25,8 @@ Requires-Dist: rapidfuzz>=3.0.0,<4
 Requires-Dist: httpx>=0.28.1
 Requires-Dist: certifi>=2025.11.12
 Requires-Dist: jinja2>=3.1.6
-Requires-Dist: unique-search-proxy-core>=2026.44.0.dev6,<2026.44.0rc0
-Requires-Dist: unique-search-proxy-sdk>=2026.44.0.dev3,<2026.44.0rc0
+Requires-Dist: unique-search-proxy-core>=2026.44.0.dev11,<2026.44.0rc0
+Requires-Dist: unique-search-proxy-sdk>=2026.44.0.dev6,<2026.44.0rc0
 Requires-Dist: openai>=1.105.0,<3
 Requires-Python: >=3.12, <4.0
 Description-Content-Type: text/markdown
@@ -73,7 +73,7 @@ Standard engines that live in proxy-core use **`ExposableParam`** (`expose` + `v
 
 ### Agent search (grounding)
 
-An upstream AI agent / grounded model searches and returns opaque text that this tool parses into `WebSearchResult`s.
+An upstream AI agent / grounded model searches and returns opaque text that this tool parses into `WebSearchResult`s. Bing results take their URLs only from the citations Foundry attaches to the answer, never from the answer text.
 
 | Engine | Proxy | Notes |
 |--------|-------|-------|

{unique_web_search-2026.44.0.dev4 → unique_web_search-2026.44.0.dev6}/README.md

@@ -40,7 +40,7 @@ Standard engines that live in proxy-core use **`ExposableParam`** (`expose` + `v
 
 ### Agent search (grounding)
 
-An upstream AI agent / grounded model searches and returns opaque text that this tool parses into `WebSearchResult`s.
+An upstream AI agent / grounded model searches and returns opaque text that this tool parses into `WebSearchResult`s. Bing results take their URLs only from the citations Foundry attaches to the answer, never from the answer text.
 
 | Engine | Proxy | Notes |
 |--------|-------|-------|

{unique_web_search-2026.44.0.dev4 → unique_web_search-2026.44.0.dev6}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "unique_web_search"
-version = "2026.44.0.dev4"
+version = "2026.44.0.dev6"
 description = ""
 readme = "README.md"
 requires-python = ">=3.12,<4.0"
@@ -22,7 +22,7 @@ dependencies = [
     "unidecode>=1.4.0,<2",
     "azure-ai-projects>=2.3,<3",
     "azure-identity>=1.25.0,<2",
-    "unique-toolkit[monitoring]>=2026.44.0.dev9,<2026.44.0rc0",
+    "unique-toolkit[monitoring]>=2026.44.0.dev10,<2026.44.0rc0",
     "azure-core>=1.36.0,<2",
     "google-cloud-aiplatform>=1.128.0,<2",
     "google-auth>=2.43.0,<3",
@@ -31,8 +31,8 @@ dependencies = [
     "httpx>=0.28.1",
     "certifi>=2025.11.12",
     "jinja2>=3.1.6",
-    "unique-search-proxy-core>=2026.44.0.dev6,<2026.44.0rc0",
-    "unique-search-proxy-sdk>=2026.44.0.dev3,<2026.44.0rc0",
+    "unique-search-proxy-core>=2026.44.0.dev11,<2026.44.0rc0",
+    "unique-search-proxy-sdk>=2026.44.0.dev6,<2026.44.0rc0",
     "openai>=1.105.0,<3",
 ]
 

{unique_web_search-2026.44.0.dev4 → unique_web_search-2026.44.0.dev6}/src/unique_web_search/metrics.py

@@ -15,6 +15,21 @@ custom_api_url_safety_report = m.counter(
     "Report-only URL safety outcomes for legacy Custom API searches",
     ["outcome", "reason_category"],
 )
+grounding_parse_failures = m.counter(
+    "grounding_parse_failures_total",
+    "Grounded agent answers that yielded no usable sources",
+    ["engine", "reason"],
+)
+grounding_dropped_citations = m.counter(
+    "grounding_dropped_citations_total",
+    "Citations dropped from grounded agent answers because they failed validation",
+    ["engine", "reason"],
+)
+grounding_dropped_lines = m.counter(
+    "grounding_dropped_lines_total",
+    "Answer lines dropped from grounded agent answers",
+    ["engine", "reason"],
+)
 
 # End-to-End Tool (P1)
 tool_duration = m.histogram(

{unique_web_search-2026.44.0.dev4 → unique_web_search-2026.44.0.dev6}/src/unique_web_search/service.py

@@ -3,7 +3,7 @@ from datetime import datetime
 from time import time
 
 from typing_extensions import override
-from unique_search_proxy_core.context import RequestContext
+from unique_search_proxy_core.context import EntryPoint, RequestContext
 from unique_search_proxy_core.param_policy.exposable_config import (
     ExposableParamsConfig,
 )
@@ -81,6 +81,7 @@ class WebSearchTool(Tool[WebSearchConfig]):
             user_id=self.event.user_id,
             chat_id=self.event.payload.chat_id,
             user_metadata=self.event.payload.user_metadata or {},
+            entry_point=EntryPoint.CHAT_TOOL,
         )
         self.search_engine_service = get_search_engine_service(
             self.config.search_engine_config,

{unique_web_search-2026.44.0.dev4 → unique_web_search-2026.44.0.dev6}/src/unique_web_search/services/proxy/bridge.py

@@ -3,11 +3,17 @@ from __future__ import annotations
 from collections.abc import AsyncIterator
 from contextlib import asynccontextmanager
 
-from unique_search_proxy_core.context import LOCAL_REQUEST_CONTEXT, RequestContext
+from unique_search_proxy_core.context import (
+    LOCAL_REQUEST_CONTEXT,
+    SERVICE_ID_HEADER,
+    RequestContext,
+)
 from unique_search_proxy_sdk import UniqueSearchProxyClient
 
 from unique_web_search.settings import env_settings
 
+SERVICE_ID = "assistants-core"
+
 search_proxy_client_enabled = env_settings.search_proxy_base_url is not None
 
 
@@ -24,6 +30,7 @@ async def open_search_proxy_client(
         timeout=timeout,
         context=context,
     ) as client:
+        client.openapi.get_async_httpx_client().headers[SERVICE_ID_HEADER] = SERVICE_ID
         yield client
 
 

{unique_web_search-2026.44.0.dev4 → unique_web_search-2026.44.0.dev6}/src/unique_web_search/services/proxy/mappers.py

@@ -3,6 +3,7 @@ from __future__ import annotations
 from collections.abc import Sequence
 from typing import TYPE_CHECKING
 
+from unique_search_proxy_core.schema import AnswerSpanCitation
 from unique_search_proxy_sdk._generated.models.agent_search_response import (
     AgentSearchResponse,
 )
@@ -71,3 +72,12 @@ def agent_answer_text(response: AgentSearchResponse) -> str:
     if isinstance(response.answer, str):
         return response.answer
     return ""
+
+
+def agent_citations(response: AgentSearchResponse) -> list[AnswerSpanCitation]:
+    if not isinstance(response.citations, list):
+        return []
+    return [
+        AnswerSpanCitation.model_validate(citation.to_dict())
+        for citation in response.citations
+    ]

{unique_web_search-2026.44.0.dev4 → unique_web_search-2026.44.0.dev6}/src/unique_web_search/services/search_engine/README.md

@@ -31,7 +31,7 @@ Deployment configs for Google / Brave / Perplexity (and agent base fields) live
 | | Standard | Agent (grounding) |
 |--|----------|-------------------|
 | Proxy route | `POST /v1/search` | `POST /v1/agent-search` |
-| Result shape | `SearchResponse` → `WebSearchResult` list | Opaque `answer` text → parsed locally |
+| Result shape | `SearchResponse` → `WebSearchResult` list | Opaque `answer` text (+ provider `citations`) → parsed locally |
 | LLM knobs | `ExposableParam` + `config.merge()` | `ExposableParam` (Bing only) |
 | Typical scrape | Often yes (Google) | Often no (`requires_scraping=false`) |
 
@@ -158,7 +158,13 @@ CUSTOM_WEB_SEARCH_API_HEADERS='{"Authorization": "Bearer ..."}'
 
 ### Bing (Grounding with Bing)
 
-**Provider:** Azure AI Foundry Agent with Bing grounding · Config extends `BingAgentConfig` with `requires_scraping`, `language_model`, etc.
+**Provider:** Azure AI Foundry Agent with Bing grounding · Config extends `BingAgentConfig` with `requires_scraping`.
+
+The agent answers with one bullet per source, each ending in a citation marker.
+`BingCitationParser` (`utils/grounding/bing/citation_parsing.py`) turns each
+cited source into one result, taking its URL and title only from Foundry's
+`url_citation` annotations. Unusable citations and lines are dropped one by one
+and counted in the `grounding_*` metrics rather than failing the whole answer.
 
 | Field | Notes |
 |-------|-------|

{unique_web_search-2026.44.0.dev4 → unique_web_search-2026.44.0.dev6}/src/unique_web_search/services/search_engine/base.py

@@ -24,6 +24,7 @@ from unique_web_search.services.proxy.bridge import (
 )
 from unique_web_search.services.proxy.mappers import (
     agent_answer_text,
+    agent_citations,
     map_agent_answer,
     map_search_response,
 )
@@ -32,6 +33,10 @@ from unique_web_search.services.search_engine.schema import (
 )
 
 if TYPE_CHECKING:
+    from collections.abc import Sequence
+
+    from unique_search_proxy_core.schema import AnswerSpanCitation
+
     from unique_web_search.services.search_engine.utils.grounding.response_parsing import (
         ResponseParser,
     )
@@ -72,7 +77,7 @@ class SearchEngine(ABC, Generic[SearchEngineConfig]):
     it, otherwise it reroutes to the local legacy implementation.
     """
 
-    # Set by agent engines (Bing/VertexAI) and consumed by the agent proxy path.
+    # Set by agent engines that use the default ``_parse_agent_answer``.
     response_parsers: list[ResponseParser]
 
     def __init__(
@@ -180,12 +185,24 @@ class SearchEngine(ABC, Generic[SearchEngineConfig]):
                 engine=engine.value,
                 **invocation,
             )
-        results = await map_agent_answer(
+        results = await self._parse_agent_answer(
             agent_answer_text(response),
-            self.response_parsers,
+            agent_citations(response),
         )
         return await self._postprocess_results(results)
 
+    async def _parse_agent_answer(
+        self,
+        answer: str,
+        citations: Sequence[AnswerSpanCitation],
+    ) -> list[WebSearchResult]:
+        """Turn an agent answer into results.
+
+        Parses only the answer text. Engines whose provider cites its sources
+        alongside the answer override this, as Bing does.
+        """
+        return await map_agent_answer(answer, self.response_parsers)
+
     def _agent_proxy_invocation(
         self,
         engine: AgentEngineType,

{unique_web_search-2026.44.0.dev4 → unique_web_search-2026.44.0.dev6}/src/unique_web_search/services/search_engine/bing.py

@@ -1,4 +1,5 @@
 import asyncio
+from collections.abc import Sequence
 from typing import Any, override
 
 from pydantic import Field
@@ -10,10 +11,13 @@ from unique_search_proxy_core.agent_engines.bing.schema import BingAgentConfig
 from unique_search_proxy_core.agent_engines.bing.settings import resolve_market
 from unique_search_proxy_core.context import LOCAL_REQUEST_CONTEXT, RequestContext
 from unique_search_proxy_core.param_policy.exposed_params import ExposedParams
-from unique_toolkit._common.default_language_model import DEFAULT_LANGUAGE_MODEL
-from unique_toolkit._common.validators import LMI, get_LMI_default_field
-from unique_toolkit.language_model import LanguageModelService
+from unique_search_proxy_core.schema import AnswerSpanCitation
 
+from unique_web_search.metrics import (
+    grounding_dropped_citations,
+    grounding_dropped_lines,
+    grounding_parse_failures,
+)
 from unique_web_search.services.proxy.bridge import (
     search_proxy_client_enabled,
 )
@@ -22,11 +26,8 @@ from unique_web_search.services.search_engine.registry import register_search_en
 from unique_web_search.services.search_engine.schema import (
     WebSearchResult,
 )
-from unique_web_search.services.search_engine.utils.grounding import (
-    JsonConversionStrategy,
-    LLMParserStrategy,
-)
 from unique_web_search.services.search_engine.utils.grounding.bing import (
+    BingCitationParser,
     create_and_process_run,
     credentials_are_valid,
     get_credentials,
@@ -35,10 +36,6 @@ from unique_web_search.services.search_engine.utils.grounding.bing import (
 
 
 class BingSearchConfig(BingAgentConfig):
-    language_model: LMI = get_LMI_default_field(
-        DEFAULT_LANGUAGE_MODEL,
-        description="The language model to use in as a fallback parser if the agent response is not a valid JSON.",
-    )
     requires_scraping: bool = Field(
         default=False,
         description="Whether the search engine requires scraping.",
@@ -51,26 +48,17 @@ class BingSearchConfig(BingAgentConfig):
     config_cls=BingSearchConfig,
     mode=SearchEngineMode.AGENT,
     config_display_name="Grounding with Bing",
-    needs_language_model=True,
 )
 class BingSearch(SearchEngine[BingSearchConfig]):
     def __init__(
         self,
         config: BingSearchConfig,
-        language_model_service: LanguageModelService,
         *,
         request_context: RequestContext = LOCAL_REQUEST_CONTEXT,
     ):
         super().__init__(config, request_context=request_context)
         self.credentials = get_credentials()
-
-        self.response_parsers = [
-            JsonConversionStrategy(),
-            LLMParserStrategy(
-                config.language_model,
-                language_model_service,
-            ),
-        ]
+        self.citation_response_parser = BingCitationParser()
 
     @property
     def is_configured(self) -> bool:
@@ -116,12 +104,33 @@ class BingSearch(SearchEngine[BingSearchConfig]):
         agent_client = get_project_client(self.credentials)
 
         async with agent_client:
-            search_results = await create_and_process_run(
+            answer, citations = await create_and_process_run(
                 agent_client,
                 query=query,
                 grounding=self._grounding_configuration(),
-                response_parsers_strategies=self.response_parsers,
                 generation_instructions=self.config.generation_instructions,
             )
 
-        return search_results
+        return await self._parse_agent_answer(answer, citations)
+
+    @override
+    async def _parse_agent_answer(
+        self,
+        answer: str,
+        citations: Sequence[AnswerSpanCitation],
+    ) -> list[WebSearchResult]:
+        parsed = await self.citation_response_parser(answer, citations)
+        engine = self.config.engine.value
+        for reason, count in parsed.dropped_citations.items():
+            grounding_dropped_citations.labels(engine=engine, reason=reason.value).inc(
+                count
+            )
+        for reason, count in parsed.dropped_lines.items():
+            grounding_dropped_lines.labels(engine=engine, reason=reason.value).inc(
+                count
+            )
+        if parsed.failure is not None:
+            grounding_parse_failures.labels(
+                engine=engine, reason=parsed.failure.value
+            ).inc()
+        return parsed.results

{unique_web_search-2026.44.0.dev4 → unique_web_search-2026.44.0.dev6}/src/unique_web_search/services/search_engine/utils/grounding/bing/__init__.py

@@ -1,3 +1,6 @@
+from unique_web_search.services.search_engine.utils.grounding.bing.citation_parsing import (
+    BingCitationParser,
+)
 from unique_web_search.services.search_engine.utils.grounding.bing.client import (
     credentials_are_valid,
     get_credentials,
@@ -8,6 +11,7 @@ from unique_web_search.services.search_engine.utils.grounding.bing.runner import
 )
 
 __all__ = [
+    "BingCitationParser",
     "credentials_are_valid",
     "get_credentials",
     "get_project_client",

unique_web_search-2026.44.0.dev6/src/unique_web_search/services/search_engine/utils/grounding/bing/citation_instructions.j2

@@ -0,0 +1,18 @@
+{{ generation_instructions }}
+
+## Source-isolated evidence list
+Use Bing Grounding to select distinct useful sources.
+The complete response must consist only of lines in this form:
+- <a paragraph of facts supported only by one source> <that source's citation>
+
+Each line is a paragraph, not a sentence. Write several sentences per line that report every relevant fact, figure, date, name and quote from that source, with as much detail as the source offers, often 100 to 200 words. Lines can be long; do not shorten them.
+
+Hard requirements:
+1. Exactly one line per distinct source and exactly one distinct source per line.
+2. Every line begins with `- ` and contains exactly one citation marker as its final content.
+3. Never synthesize, compare, or combine facts across sources within a line.
+4. Put all useful facts from a source into that source's one line; never cite it again.
+5. No headings, preface, conclusion, line breaks inside a paragraph, nested bullets, JSON, or uncited claims.
+6. Never invent, reconstruct, type, or guess citation markers or source URLs.
+
+Before responding, silently verify that the number of lines equals the number of distinct cited sources.