inspect-ai 0.3.100__py3-none-any.whl → 0.3.101__py3-none-any.whl

inspect_ai/_view/www/package.json

@@ -67,8 +67,8 @@
     "@popperjs/core": "^2.11.8",
     "ansi-output": "^0.0.9",
     "asciinema-player": "^3.9.0",
-    "bootstrap": "^5.3.3",
-    "bootstrap-icons": "^1.11.3",
+    "bootstrap": "^5.3.6",
+    "bootstrap-icons": "^1.12.1",
     "clipboard": "^2.0.11",
     "clsx": "^2.1.1",
     "codemirror": "^6.0.1",
@@ -89,8 +89,8 @@
     "react": "^19.0.0",
     "react-dom": "^19.0.0",
     "react-popper": "^2.3.0",
-    "react-router-dom": "^7.5.3",
-    "react-virtuoso": "^4.12.6",
+    "react-router-dom": "^7.6.0",
+    "react-virtuoso": "^4.12.7",
     "zustand": "^5.0.5",
     "use-resize-observer": "^9.1.0"
   }

inspect_ai/_view/www/src/app/log-view/tabs/grouping.ts

@@ -47,7 +47,7 @@ const noGrouping = (
     const itemCount = counter.item();
     return [
       {
-        label: `Sample ${itemCount}`,
+        label: `Sample ${sample.id}`,
         number: itemCount,
         index: index,
         data: sample,
@@ -107,10 +107,10 @@ const groupBySample = (
     if (sample.id !== lastId) {
       counter.incrementGroup();
       results.push({
-        label: `Sample ${itemCount}`,
+        label: `Sample ${sample.id}`,
         number: counter.group(),
         index: index,
-        data: `Sample ${counter.group()}`,
+        data: `Sample ${sample.id}`,
         type: "separator",
       } as SeparatorListItem);
       counter.resetItem();
@@ -175,7 +175,7 @@ const groupByEpoch = (
     // Compute the index within the epoch
     counter.incrementItem();
     results.push({
-      label: `Sample ${counter.item()} (Epoch ${counter.group()})`,
+      label: `Sample ${sample.id} (Epoch ${sample.epoch})`,
       number: counter.item(),
       index: index,
       data: sample,

inspect_ai/_view/www/src/app/routing/navigationHooks.ts

@@ -130,29 +130,26 @@ export const useSampleNavigation = () => {
 
   // Navigate to a specific sample with index
   const showSample = useCallback(
-    (index: number, specifiedSampleTabId?: string) => {
-      if (sampleSummaries && index >= 0 && index < sampleSummaries.length) {
-        const sample = sampleSummaries[index];
-        const resolvedPath = resolveLogPath();
-
-        if (resolvedPath) {
-          // Update internal state
-          selectSample(index);
-          setShowingSampleDialog(true);
-
-          // Use specified sampleTabId if provided, otherwise use current sampleTabId from URL params
-          const currentSampleTabId = specifiedSampleTabId || sampleTabId;
-
-          const url = sampleUrl(
-            resolvedPath,
-            sample.id,
-            sample.epoch,
-            currentSampleTabId,
-          );
-
-          // Navigate to the sample URL
-          navigate(url);
-        }
+    (
+      index: number,
+      id: string | number,
+      epoch: number,
+      specifiedSampleTabId?: string,
+    ) => {
+      const resolvedPath = resolveLogPath();
+
+      if (resolvedPath) {
+        // Update internal state
+        selectSample(index);
+        setShowingSampleDialog(true);
+
+        // Use specified sampleTabId if provided, otherwise use current sampleTabId from URL params
+        const currentSampleTabId = specifiedSampleTabId || sampleTabId;
+
+        const url = sampleUrl(resolvedPath, id, epoch, currentSampleTabId);
+
+        // Navigate to the sample URL
+        navigate(url);
       }
     },
     [
@@ -171,7 +168,7 @@ export const useSampleNavigation = () => {
     const itemsCount = sampleSummaries.length;
     const next = Math.min(selectedSampleIndex + 1, itemsCount - 1);
     if (next > -1) {
-      showSample(next, sampleTabId);
+      selectSample(next);
     }
   }, [selectedSampleIndex, showSample, sampleTabId]);
 
@@ -179,7 +176,7 @@ export const useSampleNavigation = () => {
   const previousSample = useCallback(() => {
     const prev = selectedSampleIndex - 1;
     if (prev > -1) {
-      showSample(prev, sampleTabId);
+      selectSample(prev);
     }
   }, [selectedSampleIndex, showSample, sampleTabId]);
 

inspect_ai/_view/www/src/app/samples/list/SampleList.tsx

@@ -113,11 +113,19 @@ export const SampleList: FC<SampleListProps> = memo((props) => {
           e.preventDefault();
           e.stopPropagation();
           break;
-        case "Enter":
-          sampleNavigation.showSample(selectedSampleIndex);
-          e.preventDefault();
-          e.stopPropagation();
+        case "Enter": {
+          const item = items[selectedSampleIndex];
+          if (item.type === "sample") {
+            sampleNavigation.showSample(
+              item.index,
+              item.data.id,
+              item.data.epoch,
+            );
+            e.preventDefault();
+            e.stopPropagation();
+          }
           break;
+        }
       }
     },
     [
@@ -150,7 +158,11 @@ export const SampleList: FC<SampleListProps> = memo((props) => {
               item.data.epoch,
             )}
             showSample={() => {
-              sampleNavigation.showSample(item.index);
+              sampleNavigation.showSample(
+                item.index,
+                item.data.id,
+                item.data.epoch,
+              );
             }}
           />
         );

inspect_ai/_view/www/src/state/hooks.ts

@@ -277,7 +277,7 @@ export const useCollapseSampleEvent = (
   const collapseEvent = useStore((state) => state.sampleActions.collapseEvent);
 
   return useMemo(() => {
-    const isCollapsed = collapsed !== null && collapsed[scope][id] === true;
+    const isCollapsed = collapsed !== null && collapsed[scope]?.[id] === true;
     const set = (value: boolean) => {
       log.debug("Set collapsed", id, value);
       collapseEvent(scope, id, value);

inspect_ai/_view/www/yarn.lock

@@ -2339,15 +2339,15 @@ balanced-match@^1.0.0:
   resolved "https://registry.yarnpkg.com/balanced-match/-/balanced-match-1.0.2.tgz#e83e3a7e3f300b34cb9d87f615fa0cbf357690ee"
   integrity sha512-3oSeUO0TMV67hN1AmbXsK4yaqU7tjiHlbxRDZOpH0KW9+CeX4bRAaX0Anxt0tx2MrpRpWwQaPwIlISEJhYU5Pw==
 
-bootstrap-icons@^1.11.3:
-  version "1.11.3"
-  resolved "https://registry.yarnpkg.com/bootstrap-icons/-/bootstrap-icons-1.11.3.tgz#03f9cb754ec005c52f9ee616e2e84a82cab3084b"
-  integrity sha512-+3lpHrCw/it2/7lBL15VR0HEumaBss0+f/Lb6ZvHISn1mlK83jjFpooTLsMWbIjJMDjDjOExMsTxnXSIT4k4ww==
+bootstrap-icons@^1.12.1:
+  version "1.13.1"
+  resolved "https://registry.yarnpkg.com/bootstrap-icons/-/bootstrap-icons-1.13.1.tgz#0aad3f5b55b67402990e729ce3883416f9cef6c5"
+  integrity sha512-ijombt4v6bv5CLeXvRWKy7CuM3TRTuPEuGaGKvTV5cz65rQSY8RQ2JcHt6b90cBBAC7s8fsf2EkQDldzCoXUjw==
 
-bootstrap@^5.3.3:
-  version "5.3.5"
-  resolved "https://registry.yarnpkg.com/bootstrap/-/bootstrap-5.3.5.tgz#be42cfe0d580e97ee1abb7d38ce94f5c393c9bb6"
-  integrity sha512-ct1CHKtiobRimyGzmsSldEtM03E8fcEX4Tb3dGXz1V8faRwM50+vfHwTzOxB3IlKO7m+9vTH3s/3C6T2EAPeTA==
+bootstrap@^5.3.6:
+  version "5.3.6"
+  resolved "https://registry.yarnpkg.com/bootstrap/-/bootstrap-5.3.6.tgz#fbd91ebaff093f5b191a1c01a8c866d24f9fa6e1"
+  integrity sha512-jX0GAcRzvdwISuvArXn3m7KZscWWFAf1MKBcnzaN02qWMb3jpMoUX4/qgeiGzqyIb4ojulRzs89UCUmGcFSzTA==
 
 brace-expansion@^1.1.7:
   version "1.1.11"
@@ -4478,21 +4478,20 @@ react-refresh@^0.17.0:
   resolved "https://registry.yarnpkg.com/react-refresh/-/react-refresh-0.17.0.tgz#b7e579c3657f23d04eccbe4ad2e58a8ed51e7e53"
   integrity sha512-z6F7K9bV85EfseRCp2bzrpyQ0Gkw1uLoCel9XBVWPg/TjRj94SkJzUTGfOa4bs7iJvBWtQG0Wq7wnI0syw3EBQ==
 
-react-router-dom@^7.5.3:
-  version "7.5.3"
-  resolved "https://registry.yarnpkg.com/react-router-dom/-/react-router-dom-7.5.3.tgz#496e9f6d90f731703c7772668b41747028e0a2d5"
-  integrity sha512-cK0jSaTyW4jV9SRKAItMIQfWZ/D6WEZafgHuuCb9g+SjhLolY78qc+De4w/Cz9ybjvLzShAmaIMEXt8iF1Cm+A==
+react-router-dom@^7.6.0:
+  version "7.6.1"
+  resolved "https://registry.yarnpkg.com/react-router-dom/-/react-router-dom-7.6.1.tgz#263c9102e96b58d336258a51d68080b40c28f526"
+  integrity sha512-vxU7ei//UfPYQ3iZvHuO1D/5fX3/JOqhNTbRR+WjSBWxf9bIvpWK+ftjmdfJHzPOuMQKe2fiEdG+dZX6E8uUpA==
   dependencies:
-    react-router "7.5.3"
+    react-router "7.6.1"
 
-react-router@7.5.3:
-  version "7.5.3"
-  resolved "https://registry.yarnpkg.com/react-router/-/react-router-7.5.3.tgz#9e5420832af8c3690740c1797d4fa54613fea06d"
-  integrity sha512-3iUDM4/fZCQ89SXlDa+Ph3MevBrozBAI655OAfWQlTm9nBR0IKlrmNwFow5lPHttbwvITZfkeeeZFP6zt3F7pw==
+react-router@7.6.1:
+  version "7.6.1"
+  resolved "https://registry.yarnpkg.com/react-router/-/react-router-7.6.1.tgz#a54f9b980b94594bcb4b7f26611612a9f6e17461"
+  integrity sha512-hPJXXxHJZEsPFNVbtATH7+MMX43UDeOauz+EAU4cgqTn7ojdI9qQORqS8Z0qmDlL1TclO/6jLRYUEtbWidtdHQ==
   dependencies:
     cookie "^1.0.1"
     set-cookie-parser "^2.6.0"
-    turbo-stream "2.4.0"
 
 react-transition-group@^4.4.5:
   version "4.4.5"
@@ -4504,10 +4503,10 @@ react-transition-group@^4.4.5:
     loose-envify "^1.4.0"
     prop-types "^15.6.2"
 
-react-virtuoso@^4.12.6:
-  version "4.12.6"
-  resolved "https://registry.yarnpkg.com/react-virtuoso/-/react-virtuoso-4.12.6.tgz#20fe374d43cce3c9821e29f4cc4d050596d06d01"
-  integrity sha512-bfvS6aCL1ehXmq39KRiz/vxznGUbtA27I5I24TYCe1DhMf84O3aVNCIwrSjYQjkJGJGzY46ihdN8WkYlemuhMQ==
+react-virtuoso@^4.12.7:
+  version "4.12.8"
+  resolved "https://registry.yarnpkg.com/react-virtuoso/-/react-virtuoso-4.12.8.tgz#db1dbba617f91c1dcd760aa90e09ef991e65a356"
+  integrity sha512-NMMKfDBr/+xZZqCQF3tN1SZsh6FwOJkYgThlfnsPLkaEhdyQo0EuWUzu3ix6qjnI7rYwJhMwRGoJBi+aiDfGsA==
 
 react@^19.0.0:
   version "19.1.0"
@@ -4922,11 +4921,6 @@ ts-jest@^29.3.2:
     type-fest "^4.39.1"
     yargs-parser "^21.1.1"
 
-turbo-stream@2.4.0:
-  version "2.4.0"
-  resolved "https://registry.yarnpkg.com/turbo-stream/-/turbo-stream-2.4.0.tgz#1e4fca6725e90fa14ac4adb782f2d3759a5695f0"
-  integrity sha512-FHncC10WpBd2eOmGwpmQsWLDoK4cqsA/UT/GqNoaKOQnT8uzhtCbg3EoUDMvqpOSAI0S26mr0rkjzbOO6S3v1g==
-
 type-check@^0.4.0, type-check@~0.4.0:
   version "0.4.0"
   resolved "https://registry.yarnpkg.com/type-check/-/type-check-0.4.0.tgz#07b8203bfa7056c0657050e3ccd2c37730bab8f1"

inspect_ai/analysis/beta/__init__.py

@@ -7,6 +7,7 @@ from ._dataframe.evals.columns import (
     EvalColumn,
     EvalColumns,
     EvalConfig,
+    EvalDataset,
     EvalInfo,
     EvalModel,
     EvalResults,
@@ -41,6 +42,7 @@ __all__ = [
     "EvalModel",
     "EvalColumns",
     "EvalConfig",
+    "EvalDataset",
     "EvalResults",
     "EvalScores",
     "samples_df",

inspect_ai/dataset/_sources/csv.py

@@ -14,7 +14,7 @@ from .._dataset import (
     MemoryDataset,
     RecordToSample,
 )
-from .._util import data_to_samples, record_to_sample_fn
+from .._util import data_to_samples, record_to_sample_fn, shuffle_choices_if_requested
 
 
 def csv_dataset(
@@ -88,11 +88,7 @@ def csv_dataset(
         if shuffle:
             dataset.shuffle(seed=seed)
 
-        # shuffle choices, if requested
-        if isinstance(shuffle_choices, int):
-            dataset.shuffle_choices(seed=shuffle_choices)
-        elif shuffle_choices is True:
-            dataset.shuffle_choices()
+        shuffle_choices_if_requested(dataset, shuffle_choices)
 
         # limit if requested
         if limit:

inspect_ai/dataset/_sources/hf.py

@@ -16,7 +16,7 @@ from .._dataset import (
     MemoryDataset,
     RecordToSample,
 )
-from .._util import data_to_samples, record_to_sample_fn
+from .._util import data_to_samples, record_to_sample_fn, shuffle_choices_if_requested
 
 
 def hf_dataset(
@@ -125,10 +125,6 @@ def hf_dataset(
         location=path,
     )
 
-    # maybe shuffle the choices
-    if isinstance(shuffle_choices, int):
-        memory_dataset.shuffle_choices(seed=shuffle_choices)
-    elif shuffle_choices is True:
-        memory_dataset.shuffle_choices()
+    shuffle_choices_if_requested(memory_dataset, shuffle_choices)
 
     return memory_dataset

inspect_ai/dataset/_sources/json.py

@@ -15,7 +15,7 @@ from .._dataset import (
     MemoryDataset,
     RecordToSample,
 )
-from .._util import data_to_samples, record_to_sample_fn
+from .._util import data_to_samples, record_to_sample_fn, shuffle_choices_if_requested
 from .util import resolve_sample_files
 
 
@@ -88,11 +88,7 @@ def json_dataset(
         if shuffle:
             dataset.shuffle(seed=seed)
 
-        # shuffle choices, if requested
-        if isinstance(shuffle_choices, int):
-            dataset.shuffle_choices(seed=shuffle_choices)
-        elif shuffle_choices is True:
-            dataset.shuffle_choices()
+        shuffle_choices_if_requested(dataset, shuffle_choices)
 
         # limit if requested
         if limit:

inspect_ai/dataset/_util.py

@@ -13,6 +13,7 @@ from inspect_ai.model import (
 from inspect_ai.util._sandbox.environment import SandboxEnvironmentSpec
 
 from ._dataset import (
+    Dataset,
     DatasetRecord,
     FieldSpec,
     RecordToSample,
@@ -225,3 +226,25 @@ def read_files(files: Any | None) -> dict[str, str] | None:
         raise ValueError(f"Unexpected type for 'files' field: {type(files)}")
     else:
         return None
+
+
+def shuffle_choices_if_requested(
+    dataset: Dataset, shuffle_choices: bool | int | None
+) -> None:
+    """
+    Shuffle the choices in the dataset if requested.
+
+    The `shuffle_choices` parameter passed to `json_dataset`, `csv_dataset`,
+    and `hf_dataset` can be a boolean, an integer, or `None` (default).
+    If it is a boolean, it will shuffle the choices if the value is `True`,
+    and do nothing if it is `False`.
+    If it is an integer, it will shuffle the choices using the integer as the seed.
+    """
+    # Note that `isinstance(x, int)` returns True if x is True or False,
+    # so we need to check for both explicitly
+    if shuffle_choices is True:
+        dataset.shuffle_choices()
+    elif shuffle_choices is False:
+        pass
+    elif isinstance(shuffle_choices, int):
+        dataset.shuffle_choices(seed=shuffle_choices)

inspect_ai/log/_recorders/eval.py

@@ -133,6 +133,7 @@ class EvalRecorder(FileRecorder):
         results: EvalResults | None,
         reductions: list[EvalSampleReductions] | None,
         error: EvalError | None = None,
+        header_only: bool = False,
     ) -> EvalLog:
         # get the key and log
         key = self._log_file_key(eval)
@@ -174,7 +175,7 @@ class EvalRecorder(FileRecorder):
 
         # flush and write the results
         await log.flush()
-        return await log.close()
+        return await log.close(header_only)
 
     @classmethod
     @override
@@ -321,12 +322,12 @@ class ZipLogFile:
                     # re-open zip file w/ self.temp_file pointer at end
                     self._open()
 
-    async def close(self) -> EvalLog:
+    async def close(self, header_only: bool) -> EvalLog:
         async with self._lock:
             # read the log from the temp file then close it
             try:
                 self._temp_file.seek(0)
-                return _read_log(self._temp_file, self._file)
+                return _read_log(self._temp_file, self._file, header_only=header_only)
             finally:
                 self._temp_file.close()
                 if self._zip:

inspect_ai/log/_recorders/json.py

@@ -96,6 +96,7 @@ class JSONRecorder(FileRecorder):
         results: EvalResults | None,
         reductions: list[EvalSampleReductions] | None,
         error: EvalError | None = None,
+        header_only: bool = False,
     ) -> EvalLog:
         log = self.data[self._log_file_key(spec)]
         log.data.status = status

inspect_ai/log/_recorders/recorder.py

@@ -46,6 +46,7 @@ class Recorder(abc.ABC):
         results: EvalResults | None,
         reductions: list[EvalSampleReductions] | None,
         error: EvalError | None = None,
+        header_only: bool = False,
     ) -> EvalLog: ...
 
     @classmethod

inspect_ai/model/_openai_responses.py

@@ -162,9 +162,9 @@ def openai_responses_tool_choice(
 
 
 def openai_responses_tools(
-    tools: list[ToolInfo], config: GenerateConfig
+    tools: list[ToolInfo], model_name: str, config: GenerateConfig
 ) -> list[ToolParam]:
-    return [_tool_param_for_tool_info(tool, config) for tool in tools]
+    return [_tool_param_for_tool_info(tool, model_name, config) for tool in tools]
 
 
 def openai_responses_chat_choices(
@@ -177,9 +177,11 @@ def openai_responses_chat_choices(
 
 
 def is_native_tool_configured(
-    tools: Sequence[ToolInfo], config: GenerateConfig
+    tools: Sequence[ToolInfo], model_name: str, config: GenerateConfig
 ) -> bool:
-    return any(_maybe_native_tool_param(tool, config) is not None for tool in tools)
+    return any(
+        _maybe_native_tool_param(tool, model_name, config) is not None for tool in tools
+    )
 
 
 # The next two function perform transformations between OpenAI types an Inspect
@@ -433,11 +435,13 @@ def _model_tool_call_for_internal(
 
 def _maybe_native_tool_param(
     tool: ToolInfo,
+    model_name: str,
     config: GenerateConfig,
 ) -> ToolParam | None:
     return (
         (
-            maybe_computer_use_preview_tool(tool) or maybe_web_search_tool(tool)
+            maybe_computer_use_preview_tool(tool)
+            or maybe_web_search_tool(model_name, tool)
             # or self.text_editor_tool_param(tool)
             # or self.bash_tool_param(tool)
         )
@@ -502,11 +506,12 @@ _ResponseToolCallParam = (
 
 def _tool_param_for_tool_info(
     tool: ToolInfo,
+    model_name: str,
     config: GenerateConfig,
 ) -> ToolParam:
     # Use a native tool implementation when available. Otherwise, use the
     # standard tool implementation
-    return _maybe_native_tool_param(tool, config) or FunctionToolParam(
+    return _maybe_native_tool_param(tool, model_name, config) or FunctionToolParam(
         type="function",
         name=_responses_tool_alias(tool.name),
         description=tool.description,

inspect_ai/model/_openai_web_search.py

@@ -4,11 +4,18 @@ from openai.types.responses import WebSearchTool, WebSearchToolParam
 
 from inspect_ai.tool._tool_info import ToolInfo
 
+COMPATIBLE_MODELS = ["gpt-4o", "gpt-4o-mini", "gpt-4.1"]
 
-def maybe_web_search_tool(tool: ToolInfo) -> WebSearchToolParam | None:
+
+def maybe_web_search_tool(model_name: str, tool: ToolInfo) -> WebSearchToolParam | None:
     return (
         _web_search_tool(tool.options["openai"])
-        if tool.name == "web_search" and tool.options and "openai" in tool.options
+        if (
+            tool.name == "web_search"
+            and tool.options
+            and "openai" in tool.options
+            and model_name in COMPATIBLE_MODELS
+        )
         else None
     )
 

inspect_ai/model/_providers/openai.py

@@ -242,7 +242,9 @@ class OpenAIAPI(ModelAPI):
                 tools=tools,
                 **self.completion_params(config, False),
             )
-        elif self.responses_api or is_native_tool_configured(tools, config):
+        elif self.responses_api or is_native_tool_configured(
+            tools, self.model_name, config
+        ):
             return await generate_responses(
                 client=self.client,
                 http_hooks=self._http_hooks,

inspect_ai/model/_providers/openai_responses.py

@@ -59,7 +59,11 @@ async def generate_responses(
         )
 
     # prepare request (we do this so we can log the ModelCall)
-    tool_params = openai_responses_tools(tools, config) if len(tools) > 0 else NOT_GIVEN
+    tool_params = (
+        openai_responses_tools(tools, model_name, config)
+        if len(tools) > 0
+        else NOT_GIVEN
+    )
     request = dict(
         input=await openai_responses_inputs(input, model_name, store),
         tools=tool_params,

inspect_ai/scorer/_reducer/reducer.py

@@ -121,7 +121,7 @@ def pass_at(
     def reduce(scores: list[Score]) -> Score:
         def pass_at_k(values: list[float]) -> float:
             total = len(scores)
-            correct = sum(1 for v in values if v == value)
+            correct = sum(1 for v in values if v >= value)
             if total - correct < k:
                 return 1.0
             else:

inspect_ai/tool/_tools/_web_search/_google.py

@@ -32,9 +32,10 @@ class GoogleOptions(BaseModel):
 
 
 class SearchLink:
-    def __init__(self, url: str, snippet: str) -> None:
+    def __init__(self, url: str, snippet: str, title: str) -> None:
         self.url = url
         self.snippet = snippet
+        self.title = title
 
 
 def maybe_get_google_api_keys() -> tuple[str, str] | None:
@@ -71,8 +72,7 @@ def google_search_provider(
     async def search(query: str) -> str | None:
         # limit number of concurrent searches
         page_contents: list[str] = []
-        urls: list[str] = []
-        snippets: list[str] = []
+        processed_links: list[SearchLink] = []
         search_calls = 0
 
         # Paginate through search results until we have successfully extracted num_results pages or we have reached max_provider_calls
@@ -87,8 +87,7 @@ def google_search_provider(
                         page = await page_if_relevant(link.url, query, model, client)
                         if page:
                             page_contents.append(page)
-                            urls.append(link.url)
-                            snippets.append(link.snippet)
+                            processed_links.append(link)
                     # exceptions fetching pages are very common!
                     except Exception:
                         pass
@@ -98,8 +97,18 @@ def google_search_provider(
 
             search_calls += 1
 
-        all_page_contents = "\n\n".join(page_contents)
-        return None if all_page_contents == "" else all_page_contents
+        return (
+            "\n\n".join(
+                "[{title}]({url}):\n{page_content}".format(
+                    title=link.title, url=link.url, page_content=page_content
+                )
+                for link, page_content in zip(
+                    processed_links, page_contents, strict=True
+                )
+            )
+            if processed_links
+            else None
+        )
 
     async def _search(query: str, start_idx: int) -> list[SearchLink]:
         # List of allowed parameters can be found https://developers.google.com/custom-search/v1/reference/rest/v1/cse/list
@@ -121,13 +130,21 @@ def google_search_provider(
             before_sleep=log_httpx_retry_attempt(search_url),
         )
         async def execute_search() -> httpx.Response:
+            # See https://developers.google.com/custom-search/v1/reference/rest/v1/Search
             return await client.get(search_url)
 
         result = await execute_search()
         data = result.json()
 
         if "items" in data:
-            return [SearchLink(item["link"], item["snippet"]) for item in data["items"]]
+            return [
+                SearchLink(
+                    url=item["link"],
+                    snippet=item.get("snippet", ""),  # sometimes not present
+                    title=item["title"],
+                )
+                for item in data["items"]
+            ]
         else:
             return []
 
@@ -135,13 +152,13 @@ def google_search_provider(
 
 
 async def page_if_relevant(
-    link: str, query: str, relevance_model: str | None, client: httpx.AsyncClient
+    url: str, query: str, relevance_model: str | None, client: httpx.AsyncClient
 ) -> str | None:
     """
     Use parser model to determine if a web page contents is relevant to a query.
 
     Args:
-        link (str): Web page link.
+        url (str): Web page url.
         query (str): Search query.
         relevance_model (Model): Model used to parse web pages for relevance.
         client: (httpx.Client): HTTP client to use to fetch the page
@@ -156,7 +173,7 @@ async def page_if_relevant(
 
     # retrieve document
     try:
-        response = await client.get(link)
+        response = await client.get(url)
         response.raise_for_status()
     except httpx.HTTPError as exc:
         raise Exception(f"HTTP error occurred: {exc}")

inspect_ai/tool/_tools/_web_search/_tavily.py

@@ -75,6 +75,7 @@ def tavily_search_provider(
     client = httpx.AsyncClient(timeout=30)
 
     async def search(query: str) -> str | None:
+        # See https://docs.tavily.com/documentation/api-reference/endpoint/search
         search_url = "https://api.tavily.com/search"
         headers = {
             "Authorization": f"Bearer {tavily_api_key}",
@@ -95,6 +96,15 @@ def tavily_search_provider(
             return response
 
         async with concurrency("tavily_web_search", max_connections):
-            return TavilySearchResponse.model_validate((await _search()).json()).answer
+            tavily_search_response = TavilySearchResponse.model_validate(
+                (await _search()).json()
+            )
+            results_str = "\n\n".join(
+                [
+                    f"[{result.title}]({result.url}):\n{result.content}"
+                    for result in tavily_search_response.results
+                ]
+            )
+            return f"Answer: {tavily_search_response.answer}\n\n{results_str}"
 
     return search