unique_web_search 2026.44.0.dev3__tar.gz → 2026.44.0.dev5__tar.gz

{unique_web_search-2026.44.0.dev3 → unique_web_search-2026.44.0.dev5}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.3
 Name: unique-web-search
-Version: 2026.44.0.dev3
+Version: 2026.44.0.dev5
 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.dev5,<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.dev10,<2026.44.0rc0
+Requires-Dist: unique-search-proxy-sdk>=2026.44.0.dev5,<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.dev3 → unique_web_search-2026.44.0.dev5}/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.dev3 → unique_web_search-2026.44.0.dev5}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "unique_web_search"
-version = "2026.44.0.dev3"
+version = "2026.44.0.dev5"
 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.dev5,<2026.44.0rc0",
-    "unique-search-proxy-sdk>=2026.44.0.dev3,<2026.44.0rc0",
+    "unique-search-proxy-core>=2026.44.0.dev10,<2026.44.0rc0",
+    "unique-search-proxy-sdk>=2026.44.0.dev5,<2026.44.0rc0",
     "openai>=1.105.0,<3",
 ]
 

{unique_web_search-2026.44.0.dev3 → unique_web_search-2026.44.0.dev5}/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.dev3 → unique_web_search-2026.44.0.dev5}/src/unique_web_search/services/content_processing/processing_strategies/llm_guard_judge.py

@@ -157,13 +157,9 @@ class LLMGuardJudge:
             needs_sanitization, _ = await self.judge_only(**kwargs)
             page = kwargs["page"]
             if not needs_sanitization:
-                _LOGGER.info(
-                    f"Page {page.url} does not require sanitization — keeping as-is"
-                )
+                _LOGGER.info("Page does not require sanitization — keeping as-is")
                 return page
-            _LOGGER.info(
-                f"Page {page.url} flagged — replacing content with flag message"
-            )
+            _LOGGER.info("Page flagged — replacing content with flag message")
             return page.model_copy(
                 update={
                     "content": "",
@@ -178,13 +174,9 @@ class LLMGuardJudge:
         if mode == SanitizeMode.JUDGE_THEN_SANITIZE:
             needs_sanitization, _ = await self.judge_only(**kwargs)
             if not needs_sanitization:
-                _LOGGER.info(
-                    f"Page {kwargs['page'].url} does not require sanitization — keeping as-is"
-                )
+                _LOGGER.info("Page does not require sanitization — keeping as-is")
                 return kwargs["page"]
-            _LOGGER.info(
-                f"Page {kwargs['page'].url} flagged — running full sanitize call"
-            )
+            _LOGGER.info("Page flagged — running full sanitize call")
             return await self.sanitize_page(**kwargs)
 
         raise ValueError(f"Unknown sanitize_mode: {mode}")
@@ -238,9 +230,7 @@ class LLMGuardJudge:
 
         parsed = await self._complete(JudgeResponse, system_prompt, user_prompt)
 
-        _LOGGER.info(
-            f"Judge result for {page.url}: needs_sanitization={parsed.needs_sanitization}"
-        )
+        _LOGGER.info(f"Judge result: needs_sanitization={parsed.needs_sanitization}")
 
         return parsed.needs_sanitization, page
 
@@ -272,9 +262,7 @@ class LLMGuardJudge:
             JudgeAndSanitizeResponse, system_prompt, user_prompt
         )
 
-        _LOGGER.info(
-            f"Judge result for {page.url}: needs_sanitization={parsed.needs_sanitization}"
-        )
+        _LOGGER.info(f"Judge result: needs_sanitization={parsed.needs_sanitization}")
 
         if parsed.needs_sanitization:
             page = parsed.apply_to_page(page, self._config.privacy_filter.flag_message)

{unique_web_search-2026.44.0.dev3 → unique_web_search-2026.44.0.dev5}/src/unique_web_search/services/content_processing/processing_strategies/llm_keyword_redact.py

@@ -171,7 +171,7 @@ class LLMKeywordRedact:
         )
 
         _LOGGER.info(
-            f"Running keyword extraction for page {page.url} "
+            "Running keyword extraction for page "
             f"(model={self._config.language_model.name})"
         )
 
@@ -195,7 +195,7 @@ class LLMKeywordRedact:
         )
 
         _LOGGER.info(
-            f"Extracted {len(parsed.sensitive_keywords)} sensitive keyword(s) from {page.url}"
+            f"Extracted {len(parsed.sensitive_keywords)} sensitive keyword(s) from page"
         )
 
         return parsed.apply_to_page(page)

{unique_web_search-2026.44.0.dev3 → unique_web_search-2026.44.0.dev5}/src/unique_web_search/services/content_processing/processing_strategies/llm_process.py

@@ -246,12 +246,12 @@ class LLMProcess:
 
         if not self._config.should_run(self._encoder, page.content):
             _LOGGER.warning(
-                f"Content is already short enough, skipping LLM processing for page {page.url}"
+                "Content is already short enough, skipping LLM processing for page"
             )
             return page
 
         _LOGGER.info(
-            f"Processing page {page.url} with LLM processor "
+            "Processing page with LLM processor "
             f"(sanitize={self._config.privacy_filter.sanitize}, mode={self._config.privacy_filter.sanitize_mode})"
         )
 

{unique_web_search-2026.44.0.dev3 → unique_web_search-2026.44.0.dev5}/src/unique_web_search/services/content_processing/processing_strategies/truncate.py

@@ -48,7 +48,7 @@ class Truncate:
             _LOGGER.warning("Truncate strategy is disabled, skipping")
             return page
 
-        _LOGGER.info(f"Truncating page {page.url} with truncate strategy")
+        _LOGGER.info("Truncating page with truncate strategy")
         tokens = self._encoder(page.content)
         if len(tokens) > self._config.max_tokens:
             page.content = self._decoder(tokens[: self._config.max_tokens])

{unique_web_search-2026.44.0.dev3 → unique_web_search-2026.44.0.dev5}/src/unique_web_search/services/content_processing/service.py

@@ -161,7 +161,7 @@ class ContentProcessor:
                 # Clear content to avoid overfilling the context; the chunker will
                 # render the compliance placeholder via the empty-content branch.
                 _LOGGER.error(
-                    f"Processing strategy failed for page {page.url}: {result.exception}",
+                    f"Processing strategy failed for page: {result.exception}",
                     exc_info=result.exception,
                 )
                 processed_pages.append(page.model_copy(update={"content": ""}))

{unique_web_search-2026.44.0.dev3 → unique_web_search-2026.44.0.dev5}/src/unique_web_search/services/crawlers/base.py

@@ -12,6 +12,7 @@ from unique_search_proxy_core.url_safety import (
 )
 
 from unique_web_search.metrics import crawl_blocked
+from unique_web_search.services.httpx_logging import hide_httpx_request_logs
 from unique_web_search.services.proxy.bridge import (
     open_search_proxy_client,
     search_proxy_client_enabled,
@@ -41,16 +42,17 @@ class BaseCrawler(ABC, Generic[CrawlerConfig]):
         self._request_context = request_context
 
     async def crawl(self, urls: list[str]) -> list[str]:
-        if search_proxy_client_enabled:
-            return await self._proxy_crawl(urls)
+        with hide_httpx_request_logs():
+            if search_proxy_client_enabled:
+                return await self._proxy_crawl(urls)
 
-        try:
-            targets = await UrlSafetyService.validate_batch_urls(urls)
-        except CrawlTargetValidationError as exc:
-            for target in exc.blocked_targets:
-                crawl_blocked.labels(reason_category=target.category).inc()
-            raise
-        return await self._legacy_crawl(targets)
+            try:
+                targets = await UrlSafetyService.validate_batch_urls(urls)
+            except CrawlTargetValidationError as exc:
+                for target in exc.blocked_targets:
+                    crawl_blocked.labels(reason_category=target.category).inc()
+                raise
+            return await self._legacy_crawl(targets)
 
     async def _proxy_crawl(self, urls: list[str]) -> list[str]:
         """Dump deployment config fields into the generic proxy crawl call."""

{unique_web_search-2026.44.0.dev3 → unique_web_search-2026.44.0.dev5}/src/unique_web_search/services/executors/base_executor.py

@@ -71,7 +71,7 @@ class BaseWebSearchExecutor(ABC, Generic[T]):
         self._notify_message = ""
 
         async def notify_callback() -> None:
-            _LOGGER.debug(f"{self.notify_name}: {self.notify_message}")
+            _LOGGER.debug(f"Progress update: {self.notify_name}")
             if self.tool_progress_reporter:
                 await self.tool_progress_reporter.notify_from_tool_call(
                     tool_call=tool_call,

{unique_web_search-2026.44.0.dev3 → unique_web_search-2026.44.0.dev5}/src/unique_web_search/services/executors/v1/schema.py

@@ -16,6 +16,7 @@ class WebSearchToolParameters(ExposedParams):
         alias_generator=to_camel,
         populate_by_name=True,
         extra="forbid",
+        hide_input_in_errors=True,
     )
 
     query: str = Field(description=_DEFAULT_QUERY_DESCRIPTION)

{unique_web_search-2026.44.0.dev3 → unique_web_search-2026.44.0.dev5}/src/unique_web_search/services/executors/v2/executor.py

@@ -212,7 +212,7 @@ class WebSearchV2Executor(BaseWebSearchExecutor[WebSearchPlan]):
         )
         delta_time = time() - time_start
         _LOGGER.debug(
-            f"Crawled {step.query_or_url} with {self.crawler_service} completed in {delta_time} seconds"
+            f"Crawled URL with {self.crawler_service} completed in {delta_time} seconds"
         )
 
         self.debug_info.steps.append(

{unique_web_search-2026.44.0.dev3 → unique_web_search-2026.44.0.dev5}/src/unique_web_search/services/executors/v2/schema.py

@@ -26,7 +26,7 @@ class Step(BaseModel):
 
 
 class WebSearchPlan(BaseModel):
-    model_config = ConfigDict(extra="forbid")
+    model_config = ConfigDict(extra="forbid", hide_input_in_errors=True)
     objective: str = Field(description="The objective of the plan")
     query_analysis: str = Field(
         description="Analysis of the user's query and what information is needed"

{unique_web_search-2026.44.0.dev3 → unique_web_search-2026.44.0.dev5}/src/unique_web_search/services/executors/v3/schema.py

@@ -15,6 +15,7 @@ _LLM_TOOL_MODEL_CONFIG = ConfigDict(
     alias_generator=to_camel,
     populate_by_name=True,
     extra="forbid",
+    hide_input_in_errors=True,
 )
 
 

unique_web_search-2026.44.0.dev5/src/unique_web_search/services/httpx_logging.py

@@ -0,0 +1,56 @@
+"""Hide httpx request logs while a web search call runs.
+
+httpx logs every request URL at INFO, and web search request URLs carry
+search queries, crawled URLs and provider credentials. httpx logs from other
+code in the same process are kept.
+"""
+
+from __future__ import annotations
+
+import logging
+from collections.abc import Iterator
+from contextlib import contextmanager
+from contextvars import ContextVar
+
+from unique_search_proxy_core.logging import HttpxRequestLogFilter
+
+_hide_httpx_request_logs: ContextVar[bool] = ContextVar(
+    "unique_web_search_hide_httpx_request_logs", default=False
+)
+
+
+# Wraps instead of subclassing HttpxRequestLogFilter: the SDK's global
+# suppress_httpx_request_logs() skips installing itself when it finds an
+# instance of that class on the httpx logger.
+class _WebSearchHttpxRequestLogFilter(logging.Filter):
+    _request_log_filter = HttpxRequestLogFilter()
+
+    def filter(self, record: logging.LogRecord) -> bool:
+        if not _hide_httpx_request_logs.get():
+            return True
+        return self._request_log_filter.filter(record)
+
+
+def _install_filter() -> None:
+    httpx_logger = logging.getLogger("httpx")
+    if any(
+        isinstance(log_filter, _WebSearchHttpxRequestLogFilter)
+        for log_filter in httpx_logger.filters
+    ):
+        return
+    httpx_logger.addFilter(_WebSearchHttpxRequestLogFilter())
+
+
+@contextmanager
+def hide_httpx_request_logs() -> Iterator[None]:
+    """Drop httpx request logs from the current task and tasks it starts.
+
+    Scoping follows ``contextvars``: work handed to threads without a copied
+    context (e.g. ``loop.run_in_executor``) is not covered.
+    """
+    _install_filter()
+    token = _hide_httpx_request_logs.set(True)
+    try:
+        yield
+    finally:
+        _hide_httpx_request_logs.reset(token)

{unique_web_search-2026.44.0.dev3 → unique_web_search-2026.44.0.dev5}/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.dev3 → unique_web_search-2026.44.0.dev5}/src/unique_web_search/services/query_elicitation.py

@@ -50,7 +50,7 @@ class QueryElicitationModel(BaseModel):
     creation with default values for form pre-population.
     """
 
-    model_config = ConfigDict(title="Query Elicitation")
+    model_config = ConfigDict(title="Query Elicitation", hide_input_in_errors=True)
     queries: list[str] = Field(description="The queries to search the web for")
 
     @classmethod
@@ -160,13 +160,13 @@ class QueryElicitationService:
             elif elicitation.status == ElicitationStatus.DECLINED:
                 _LOGGER.info(f"Query elicitation {elicitation.id} declined")
                 raise ElicitationDeclinedException(
-                    f"Elicitation triggered with queries {queries} was declined"
+                    f"Query elicitation {elicitation.id} was declined"
                 )
 
             elif elicitation.status == ElicitationStatus.CANCELLED:
                 _LOGGER.info(f"Query elicitation {elicitation.id} cancelled")
                 raise ElicitationCancelledException(
-                    f"Elicitation triggered with queries {queries} was cancelled"
+                    f"Query elicitation {elicitation.id} was cancelled"
                 )
 
         raise ElicitationExpiredException(

{unique_web_search-2026.44.0.dev3 → unique_web_search-2026.44.0.dev5}/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.dev3 → unique_web_search-2026.44.0.dev5}/src/unique_web_search/services/search_engine/base.py

@@ -17,12 +17,14 @@ from unique_search_proxy_core.param_policy.exposable_param import (
 from unique_search_proxy_core.param_policy.exposed_params import ExposedParams
 from unique_search_proxy_core.search_engines.base import SearchEngineType
 
+from unique_web_search.services.httpx_logging import hide_httpx_request_logs
 from unique_web_search.services.proxy.bridge import (
     open_search_proxy_client,
     search_proxy_client_enabled,
 )
 from unique_web_search.services.proxy.mappers import (
     agent_answer_text,
+    agent_citations,
     map_agent_answer,
     map_search_response,
 )
@@ -31,6 +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,
     )
@@ -71,7 +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__(
@@ -110,14 +116,15 @@ class SearchEngine(ABC, Generic[SearchEngineConfig]):
         params: ExposedParams | None = None,
     ) -> list[WebSearchResult]:
         """Search the web for the given query using the search engine."""
-        proxy_engine = self._proxy_engine
-        if (
-            search_proxy_client_enabled
-            and proxy_engine is not None
-            and await self._proxy_routing_enabled()
-        ):
-            return await self._proxy_search(query, params, proxy_engine)
-        return await self._legacy_search(query=query, params=params)
+        with hide_httpx_request_logs():
+            proxy_engine = self._proxy_engine
+            if (
+                search_proxy_client_enabled
+                and proxy_engine is not None
+                and await self._proxy_routing_enabled()
+            ):
+                return await self._proxy_search(query, params, proxy_engine)
+            return await self._legacy_search(query=query, params=params)
 
     async def _proxy_search(
         self,
@@ -178,12 +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.dev3 → unique_web_search-2026.44.0.dev5}/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.dev3 → unique_web_search-2026.44.0.dev5}/src/unique_web_search/services/search_engine/custom_api.py

@@ -191,7 +191,7 @@ class CustomAPI(SearchEngine[CustomAPIConfig]):
 
         if not response.is_success:
             raise ValueError(
-                f"Search engine request failed with status {response.status_code}: {response.text}"
+                f"Search engine request failed with status {response.status_code}"
             )
 
         validated_response = WebSearchResults.model_validate(response.json())

{unique_web_search-2026.44.0.dev3 → unique_web_search-2026.44.0.dev5}/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",