arize-phoenix 5.10.0__tar.gz → 5.12.0__tar.gz

{arize_phoenix-5.10.0 → arize_phoenix-5.12.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.3
 Name: arize-phoenix
-Version: 5.10.0
+Version: 5.12.0
 Summary: AI Observability and Evaluation
 Project-URL: Documentation, https://docs.arize.com/phoenix/
 Project-URL: Issues, https://github.com/Arize-ai/phoenix/issues
@@ -53,6 +53,7 @@ Requires-Dist: wrapt
 Provides-Extra: container
 Requires-Dist: anthropic; extra == 'container'
 Requires-Dist: fast-hdbscan>=0.2.0; extra == 'container'
+Requires-Dist: google-generativeai; extra == 'container'
 Requires-Dist: numba>=0.60.0; extra == 'container'
 Requires-Dist: openai>=1.0.0; extra == 'container'
 Requires-Dist: opentelemetry-exporter-otlp; extra == 'container'

{arize_phoenix-5.10.0 → arize_phoenix-5.12.0}/pyproject.toml

@@ -121,6 +121,7 @@ pg = [
 ]
 container = [
   "anthropic",
+  "google-generativeai",
   "prometheus-client",
   "openai>=1.0.0",
   "opentelemetry-sdk",

{arize_phoenix-5.10.0 → arize_phoenix-5.12.0}/src/phoenix/config.py

@@ -97,6 +97,13 @@ ENV_PHOENIX_SERVER_INSTRUMENTATION_OTLP_TRACE_COLLECTOR_GRPC_ENDPOINT = (
 ENV_PHOENIX_ENABLE_AUTH = "PHOENIX_ENABLE_AUTH"
 ENV_PHOENIX_DISABLE_RATE_LIMIT = "PHOENIX_DISABLE_RATE_LIMIT"
 ENV_PHOENIX_SECRET = "PHOENIX_SECRET"
+ENV_PHOENIX_DEFAULT_ADMIN_INITIAL_PASSWORD = "PHOENIX_DEFAULT_ADMIN_INITIAL_PASSWORD"
+"""
+The initial password for the default admin account, which defaults to ‘admin’ if not
+explicitly set. Note that changing this value will have no effect if the default admin
+record already exists in the database. In such cases, the default admin password must
+be updated manually in the application.
+"""
 ENV_PHOENIX_API_KEY = "PHOENIX_API_KEY"
 ENV_PHOENIX_USE_SECURE_COOKIES = "PHOENIX_USE_SECURE_COOKIES"
 ENV_PHOENIX_ACCESS_TOKEN_EXPIRY_MINUTES = "PHOENIX_ACCESS_TOKEN_EXPIRY_MINUTES"
@@ -274,6 +281,12 @@ def get_env_phoenix_secret() -> Optional[str]:
     return phoenix_secret
 
 
+def get_env_default_admin_initial_password() -> str:
+    from phoenix.auth import DEFAULT_ADMIN_PASSWORD
+
+    return os.environ.get(ENV_PHOENIX_DEFAULT_ADMIN_INITIAL_PASSWORD) or DEFAULT_ADMIN_PASSWORD
+
+
 def get_env_phoenix_use_secure_cookies() -> bool:
     return _bool_val(ENV_PHOENIX_USE_SECURE_COOKIES, False)
 

{arize_phoenix-5.10.0 → arize_phoenix-5.12.0}/src/phoenix/db/facilitator.py

@@ -13,13 +13,13 @@ from sqlalchemy.ext.asyncio import AsyncSession
 
 from phoenix.auth import (
     DEFAULT_ADMIN_EMAIL,
-    DEFAULT_ADMIN_PASSWORD,
     DEFAULT_ADMIN_USERNAME,
     DEFAULT_SECRET_LENGTH,
     DEFAULT_SYSTEM_EMAIL,
     DEFAULT_SYSTEM_USERNAME,
     compute_password_hash,
 )
+from phoenix.config import get_env_default_admin_initial_password
 from phoenix.db import models
 from phoenix.db.enums import COLUMN_ENUMS, UserRole
 from phoenix.server.types import DbSessionFactory
@@ -97,7 +97,8 @@ async def _ensure_user_roles(session: AsyncSession) -> None:
         admin_role_id := role_ids.get(admin_role)
     ) is not None:
         salt = secrets.token_bytes(DEFAULT_SECRET_LENGTH)
-        compute = partial(compute_password_hash, password=DEFAULT_ADMIN_PASSWORD, salt=salt)
+        password = get_env_default_admin_initial_password()
+        compute = partial(compute_password_hash, password=password, salt=salt)
         loop = asyncio.get_running_loop()
         hash_ = await loop.run_in_executor(None, compute)
         admin_user = models.User(

{arize_phoenix-5.10.0 → arize_phoenix-5.12.0}/src/phoenix/server/api/helpers/playground_clients.py

@@ -2,6 +2,7 @@ import asyncio
 import importlib.util
 import inspect
 import json
+import os
 import time
 from abc import ABC, abstractmethod
 from collections.abc import AsyncIterator, Callable, Iterator
@@ -25,6 +26,7 @@ from phoenix.evals.models.rate_limiters import (
     RateLimiter,
     RateLimitError,
 )
+from phoenix.server.api.exceptions import BadRequest
 from phoenix.server.api.helpers.playground_registry import PROVIDER_DEFAULT, register_llm_client
 from phoenix.server.api.input_types.GenerativeModelInput import GenerativeModelInput
 from phoenix.server.api.input_types.InvocationParameters import (
@@ -99,9 +101,9 @@ class PlaygroundRateLimiter(RateLimiter, KeyedSingleton):
         super().__init__(
             rate_limit_error=rate_limit_error,
             max_rate_limit_retries=3,
-            initial_per_second_request_rate=2.0,
-            maximum_per_second_request_rate=10.0,
-            enforcement_window_minutes=1,
+            initial_per_second_request_rate=1.0,
+            maximum_per_second_request_rate=3.0,
+            enforcement_window_minutes=0.05,
             rate_reduction_factor=0.5,
             rate_increase_factor=0.01,
             cooldown_seconds=5,
@@ -128,10 +130,11 @@ class PlaygroundRateLimiter(RateLimiter, KeyedSingleton):
                     self._rate_limit_handling.set()  # Set the event as a failsafe
                 await self._throttler.async_wait_until_ready()
                 request_start_time = time.time()
-                if inspect.iscoroutinefunction(fn):
-                    return await fn(*args, **kwargs)  # type: ignore
+                maybe_coroutine = fn(*args, **kwargs)
+                if inspect.isawaitable(maybe_coroutine):
+                    return await maybe_coroutine  # type: ignore[no-any-return]
                 else:
-                    return fn(*args, **kwargs)
+                    return maybe_coroutine
             except self._rate_limit_error:
                 async with self._rate_limit_handling_lock:
                     self._rate_limit_handling.clear()  # prevent new requests from starting
@@ -141,10 +144,11 @@ class PlaygroundRateLimiter(RateLimiter, KeyedSingleton):
                             try:
                                 request_start_time = time.time()
                                 await self._throttler.async_wait_until_ready()
-                                if inspect.iscoroutinefunction(fn):
-                                    return await fn(*args, **kwargs)  # type: ignore
+                                maybe_coroutine = fn(*args, **kwargs)
+                                if inspect.isawaitable(maybe_coroutine):
+                                    return await maybe_coroutine  # type: ignore[no-any-return]
                                 else:
-                                    return fn(*args, **kwargs)
+                                    return maybe_coroutine
                             except self._rate_limit_error:
                                 self._throttler.on_rate_limit_error(
                                     request_start_time, verbose=self._verbose
@@ -258,6 +262,10 @@ class OpenAIStreamingClient(PlaygroundStreamingClient):
         from openai import AsyncOpenAI
         from openai import RateLimitError as OpenAIRateLimitError
 
+        # todo: check if custom base url is set before raising error to allow
+        # for custom endpoints that don't require an API key
+        if not (api_key := api_key or os.environ.get("OPENAI_API_KEY")):
+            raise BadRequest("An API key is required for OpenAI models")
         super().__init__(model=model, api_key=api_key)
         self._attributes[LLM_PROVIDER] = OpenInferenceLLMProviderValues.OPENAI.value
         self._attributes[LLM_SYSTEM] = OpenInferenceLLMSystemValues.OPENAI.value
@@ -288,12 +296,14 @@ class OpenAIStreamingClient(PlaygroundStreamingClient):
             BoundedFloatInvocationParameter(
                 invocation_name="frequency_penalty",
                 label="Frequency Penalty",
+                default_value=0.0,
                 min_value=-2.0,
                 max_value=2.0,
             ),
             BoundedFloatInvocationParameter(
                 invocation_name="presence_penalty",
                 label="Presence Penalty",
+                default_value=0.0,
                 min_value=-2.0,
                 max_value=2.0,
             ),
@@ -306,12 +316,14 @@ class OpenAIStreamingClient(PlaygroundStreamingClient):
                 invocation_name="top_p",
                 canonical_name=CanonicalParameterName.TOP_P,
                 label="Top P",
+                default_value=1.0,
                 min_value=0.0,
                 max_value=1.0,
             ),
             IntInvocationParameter(
                 invocation_name="seed",
                 canonical_name=CanonicalParameterName.RANDOM_SEED,
+                default_value=0,
                 label="Seed",
             ),
             JSONInvocationParameter(
@@ -338,10 +350,14 @@ class OpenAIStreamingClient(PlaygroundStreamingClient):
         from openai.types.chat import ChatCompletionStreamOptionsParam
 
         # Convert standard messages to OpenAI messages
-        openai_messages = [self.to_openai_chat_completion_param(*message) for message in messages]
+        openai_messages = []
+        for message in messages:
+            openai_message = self.to_openai_chat_completion_param(*message)
+            if openai_message is not None:
+                openai_messages.append(openai_message)
         tool_call_ids: dict[int, str] = {}
         token_usage: Optional["CompletionUsage"] = None
-        throttled_create = self.rate_limiter.alimit(self.client.chat.completions.create)
+        throttled_create = self.rate_limiter._alimit(self.client.chat.completions.create)
         async for chunk in await throttled_create(
             messages=openai_messages,
             model=self.model_name,
@@ -388,7 +404,7 @@ class OpenAIStreamingClient(PlaygroundStreamingClient):
         content: JSONScalarType,
         tool_call_id: Optional[str] = None,
         tool_calls: Optional[list[JSONScalarType]] = None,
-    ) -> "ChatCompletionMessageParam":
+    ) -> Optional["ChatCompletionMessageParam"]:
         from openai.types.chat import (
             ChatCompletionAssistantMessageParam,
             ChatCompletionSystemMessageParam,
@@ -480,6 +496,7 @@ class OpenAIO1StreamingClient(OpenAIStreamingClient):
                 invocation_name="seed",
                 canonical_name=CanonicalParameterName.RANDOM_SEED,
                 label="Seed",
+                default_value=0,
             ),
             JSONInvocationParameter(
                 invocation_name="tool_choice",
@@ -488,65 +505,7 @@ class OpenAIO1StreamingClient(OpenAIStreamingClient):
             ),
         ]
 
-    async def chat_completion_create(
-        self,
-        messages: list[
-            tuple[ChatCompletionMessageRole, str, Optional[str], Optional[list[JSONScalarType]]]
-        ],
-        tools: list[JSONScalarType],
-        **invocation_parameters: Any,
-    ) -> AsyncIterator[ChatCompletionChunk]:
-        from openai import NOT_GIVEN
-
-        # Convert standard messages to OpenAI messages
-        unfiltered_openai_messages = [
-            self.to_openai_o1_chat_completion_param(*message) for message in messages
-        ]
-
-        # filter out unsupported messages
-        openai_messages: list[ChatCompletionMessageParam] = [
-            message for message in unfiltered_openai_messages if message is not None
-        ]
-
-        tool_call_ids: dict[int, str] = {}
-
-        throttled_create = self.rate_limiter.alimit(self.client.chat.completions.create)
-        response = await throttled_create(
-            messages=openai_messages,
-            model=self.model_name,
-            tools=tools or NOT_GIVEN,
-            **invocation_parameters,
-        )
-
-        choice = response.choices[0]
-        message = choice.message
-        content = message.content
-
-        text_chunk = TextChunk(content=content)
-        yield text_chunk
-
-        if (tool_calls := message.tool_calls) is not None:
-            for tool_call_index, tool_call in enumerate(tool_calls):
-                tool_call_id = (
-                    tool_call.id
-                    if tool_call.id is not None
-                    else tool_call_ids.get(tool_call_index, f"tool_call_{tool_call_index}")
-                )
-                tool_call_ids[tool_call_index] = tool_call_id
-                if (function := tool_call.function) is not None:
-                    tool_call_chunk = ToolCallChunk(
-                        id=tool_call_id,
-                        function=FunctionCallChunk(
-                            name=function.name or "",
-                            arguments=function.arguments or "",
-                        ),
-                    )
-                    yield tool_call_chunk
-
-        if (usage := response.usage) is not None:
-            self._attributes.update(dict(self._llm_token_counts(usage)))
-
-    def to_openai_o1_chat_completion_param(
+    def to_openai_chat_completion_param(
         self,
         role: ChatCompletionMessageRole,
         content: JSONScalarType,
@@ -618,12 +577,14 @@ class AzureOpenAIStreamingClient(OpenAIStreamingClient):
         super().__init__(model=model, api_key=api_key)
         self._attributes[LLM_PROVIDER] = OpenInferenceLLMProviderValues.AZURE.value
         self._attributes[LLM_SYSTEM] = OpenInferenceLLMSystemValues.OPENAI.value
-        if model.endpoint is None or model.api_version is None:
-            raise ValueError("endpoint and api_version are required for Azure OpenAI models")
+        if not (endpoint := model.endpoint or os.environ.get("AZURE_OPENAI_ENDPOINT")):
+            raise BadRequest("An Azure endpoint is required for Azure OpenAI models")
+        if not (api_version := model.api_version or os.environ.get("OPENAI_API_VERSION")):
+            raise BadRequest("An OpenAI API version is required for Azure OpenAI models")
         self.client = AsyncAzureOpenAI(
             api_key=api_key,
-            azure_endpoint=model.endpoint,
-            api_version=model.api_version,
+            azure_endpoint=endpoint,
+            api_version=api_version,
         )
 
 
@@ -631,8 +592,12 @@ class AzureOpenAIStreamingClient(OpenAIStreamingClient):
     provider_key=GenerativeProviderKey.ANTHROPIC,
     model_names=[
         PROVIDER_DEFAULT,
+        "claude-3-5-sonnet-latest",
+        "claude-3-5-haiku-latest",
+        "claude-3-5-sonnet-20241022",
+        "claude-3-5-haiku-20241022",
         "claude-3-5-sonnet-20240620",
-        "claude-3-opus-20240229",
+        "claude-3-opus-latest",
         "claude-3-sonnet-20240229",
         "claude-3-haiku-20240307",
     ],
@@ -648,6 +613,8 @@ class AnthropicStreamingClient(PlaygroundStreamingClient):
         super().__init__(model=model, api_key=api_key)
         self._attributes[LLM_PROVIDER] = OpenInferenceLLMProviderValues.ANTHROPIC.value
         self._attributes[LLM_SYSTEM] = OpenInferenceLLMSystemValues.ANTHROPIC.value
+        if not (api_key := api_key or os.environ.get("ANTHROPIC_API_KEY")):
+            raise BadRequest("An API key is required for Anthropic models")
         self.client = anthropic.AsyncAnthropic(api_key=api_key)
         self.model_name = model.name
         self.rate_limiter = PlaygroundRateLimiter(model.provider_key, anthropic.RateLimitError)
@@ -663,12 +630,14 @@ class AnthropicStreamingClient(PlaygroundStreamingClient):
                 invocation_name="max_tokens",
                 canonical_name=CanonicalParameterName.MAX_COMPLETION_TOKENS,
                 label="Max Tokens",
+                default_value=1024,
                 required=True,
             ),
             BoundedFloatInvocationParameter(
                 invocation_name="temperature",
                 canonical_name=CanonicalParameterName.TEMPERATURE,
                 label="Temperature",
+                default_value=0.0,
                 min_value=0.0,
                 max_value=1.0,
             ),
@@ -681,6 +650,7 @@ class AnthropicStreamingClient(PlaygroundStreamingClient):
                 invocation_name="top_p",
                 canonical_name=CanonicalParameterName.TOP_P,
                 label="Top P",
+                default_value=1.0,
                 min_value=0.0,
                 max_value=1.0,
             ),
@@ -707,7 +677,6 @@ class AnthropicStreamingClient(PlaygroundStreamingClient):
             "messages": anthropic_messages,
             "model": self.model_name,
             "system": system_prompt,
-            "max_tokens": 1024,
             "tools": tools,
             **invocation_parameters,
         }
@@ -820,6 +789,12 @@ class GeminiStreamingClient(PlaygroundStreamingClient):
         super().__init__(model=model, api_key=api_key)
         self._attributes[LLM_PROVIDER] = OpenInferenceLLMProviderValues.GOOGLE.value
         self._attributes[LLM_SYSTEM] = OpenInferenceLLMSystemValues.VERTEXAI.value
+        if not (
+            api_key := api_key
+            or os.environ.get("GEMINI_API_KEY")
+            or os.environ.get("GOOGLE_API_KEY")
+        ):
+            raise BadRequest("An API key is required for Gemini models")
         google_genai.configure(api_key=api_key)
         self.model_name = model.name
 
@@ -851,21 +826,25 @@ class GeminiStreamingClient(PlaygroundStreamingClient):
             FloatInvocationParameter(
                 invocation_name="presence_penalty",
                 label="Presence Penalty",
+                default_value=0.0,
             ),
             FloatInvocationParameter(
                 invocation_name="frequency_penalty",
                 label="Frequency Penalty",
+                default_value=0.0,
             ),
             BoundedFloatInvocationParameter(
                 invocation_name="top_p",
                 canonical_name=CanonicalParameterName.TOP_P,
                 label="Top P",
+                default_value=1.0,
                 min_value=0.0,
                 max_value=1.0,
             ),
             BoundedFloatInvocationParameter(
                 invocation_name="top_k",
                 label="Top K",
+                default_value=1.0,
                 min_value=0.0,
                 max_value=1.0,
             ),
@@ -873,6 +852,7 @@ class GeminiStreamingClient(PlaygroundStreamingClient):
                 invocation_name="seed",
                 canonical_name=CanonicalParameterName.RANDOM_SEED,
                 label="Seed",
+                default_value=0,
             ),
         ]
 
@@ -907,6 +887,13 @@ class GeminiStreamingClient(PlaygroundStreamingClient):
         chat = client.start_chat(history=gemini_message_history)
         stream = await chat.send_message_async(**gemini_params)
         async for event in stream:
+            self._attributes.update(
+                {
+                    LLM_TOKEN_COUNT_PROMPT: event.usage_metadata.prompt_token_count,
+                    LLM_TOKEN_COUNT_COMPLETION: event.usage_metadata.candidates_token_count,
+                    LLM_TOKEN_COUNT_TOTAL: event.usage_metadata.total_token_count,
+                }
+            )
             yield TextChunk(content=event.text)
 
     def _build_gemini_messages(

{arize_phoenix-5.10.0 → arize_phoenix-5.12.0}/src/phoenix/server/api/helpers/playground_spans.py

@@ -379,6 +379,11 @@ def _llm_output_messages(
     if content := "".join(chunk.content for chunk in text_chunks):
         yield f"{LLM_OUTPUT_MESSAGES}.0.{MESSAGE_CONTENT}", content
     for tool_call_index, (_tool_call_id, tool_call_chunks_) in enumerate(tool_call_chunks.items()):
+        if _tool_call_id:
+            yield (
+                f"{LLM_OUTPUT_MESSAGES}.0.{MESSAGE_TOOL_CALLS}.{tool_call_index}.{TOOL_CALL_ID}",
+                _tool_call_id,
+            )
         if tool_call_chunks_ and (name := tool_call_chunks_[0].function.name):
             yield (
                 f"{LLM_OUTPUT_MESSAGES}.0.{MESSAGE_TOOL_CALLS}.{tool_call_index}.{TOOL_CALL_FUNCTION_NAME}",
@@ -441,6 +446,7 @@ MESSAGE_CONTENT = MessageAttributes.MESSAGE_CONTENT
 MESSAGE_ROLE = MessageAttributes.MESSAGE_ROLE
 MESSAGE_TOOL_CALLS = MessageAttributes.MESSAGE_TOOL_CALLS
 
+TOOL_CALL_ID = ToolCallAttributes.TOOL_CALL_ID
 TOOL_CALL_FUNCTION_NAME = ToolCallAttributes.TOOL_CALL_FUNCTION_NAME
 TOOL_CALL_FUNCTION_ARGUMENTS_JSON = ToolCallAttributes.TOOL_CALL_FUNCTION_ARGUMENTS_JSON
 TOOL_CALL_ID = ToolCallAttributes.TOOL_CALL_ID

{arize_phoenix-5.10.0 → arize_phoenix-5.12.0}/src/phoenix/server/api/mutations/chat_mutations.py

@@ -1,9 +1,9 @@
 import asyncio
 from dataclasses import asdict, field
 from datetime import datetime, timezone
-from itertools import chain
+from itertools import chain, islice
 from traceback import format_exc
-from typing import Any, Iterable, Iterator, List, Optional, Union
+from typing import Any, Iterable, Iterator, List, Optional, TypeVar, Union
 
 import strawberry
 from openinference.instrumentation import safe_json_dumps
@@ -25,8 +25,9 @@ from typing_extensions import assert_never
 from phoenix.datetime_utils import local_now, normalize_datetime
 from phoenix.db import models
 from phoenix.db.helpers import get_dataset_example_revisions
+from phoenix.server.api.auth import IsNotReadOnly
 from phoenix.server.api.context import Context
-from phoenix.server.api.exceptions import BadRequest, NotFound
+from phoenix.server.api.exceptions import BadRequest, CustomGraphQLError, NotFound
 from phoenix.server.api.helpers.playground_clients import (
     PlaygroundStreamingClient,
     initialize_playground_clients,
@@ -67,6 +68,7 @@ from phoenix.utilities.json import jsonify
 from phoenix.utilities.template_formatters import (
     FStringTemplateFormatter,
     MustacheTemplateFormatter,
+    NoOpFormatter,
     TemplateFormatter,
 )
 
@@ -117,7 +119,7 @@ class ChatCompletionOverDatasetMutationPayload:
 
 @strawberry.type
 class ChatCompletionMutationMixin:
-    @strawberry.mutation
+    @strawberry.mutation(permission_classes=[IsNotReadOnly])  # type: ignore
     @classmethod
     async def chat_completion_over_dataset(
         cls,
@@ -127,11 +129,19 @@ class ChatCompletionMutationMixin:
         provider_key = input.model.provider_key
         llm_client_class = PLAYGROUND_CLIENT_REGISTRY.get_client(provider_key, input.model.name)
         if llm_client_class is None:
-            raise BadRequest(f"No LLM client registered for provider '{provider_key}'")
-        llm_client = llm_client_class(
-            model=input.model,
-            api_key=input.api_key,
-        )
+            raise BadRequest(f"Unknown LLM provider: '{provider_key.value}'")
+        try:
+            llm_client = llm_client_class(
+                model=input.model,
+                api_key=input.api_key,
+            )
+        except CustomGraphQLError:
+            raise
+        except Exception as error:
+            raise BadRequest(
+                f"Failed to connect to LLM API for {provider_key.value} {input.model.name}: "
+                f"{str(error)}"
+            )
         dataset_id = from_global_id_with_expected_type(input.dataset_id, Dataset.__name__)
         dataset_version_id = (
             from_global_id_with_expected_type(
@@ -158,7 +168,9 @@ class ChatCompletionMutationMixin:
             revisions = [
                 revision
                 async for revision in await session.stream_scalars(
-                    get_dataset_example_revisions(resolved_version_id)
+                    get_dataset_example_revisions(resolved_version_id).order_by(
+                        models.DatasetExampleRevision.id
+                    )
                 )
             ]
             if not revisions:
@@ -181,28 +193,32 @@ class ChatCompletionMutationMixin:
             session.add(experiment)
             await session.flush()
 
+        results = []
+        batch_size = 3
         start_time = datetime.now(timezone.utc)
-        results = await asyncio.gather(
-            *(
-                cls._chat_completion(
-                    info,
-                    llm_client,
-                    ChatCompletionInput(
-                        model=input.model,
-                        api_key=input.api_key,
-                        messages=input.messages,
-                        tools=input.tools,
-                        invocation_parameters=input.invocation_parameters,
-                        template=TemplateOptions(
-                            language=input.template_language,
-                            variables=revision.input,
+        for batch in _get_batches(revisions, batch_size):
+            batch_results = await asyncio.gather(
+                *(
+                    cls._chat_completion(
+                        info,
+                        llm_client,
+                        ChatCompletionInput(
+                            model=input.model,
+                            api_key=input.api_key,
+                            messages=input.messages,
+                            tools=input.tools,
+                            invocation_parameters=input.invocation_parameters,
+                            template=TemplateOptions(
+                                language=input.template_language,
+                                variables=revision.input,
+                            ),
                         ),
-                    ),
-                )
-                for revision in revisions
-            ),
-            return_exceptions=True,
-        )
+                    )
+                    for revision in batch
+                ),
+                return_exceptions=True,
+            )
+            results.extend(batch_results)
 
         payload = ChatCompletionOverDatasetMutationPayload(
             dataset_id=GlobalID(models.Dataset.__name__, str(dataset.id)),
@@ -258,7 +274,7 @@ class ChatCompletionMutationMixin:
             payload.examples.append(example_payload)
         return payload
 
-    @strawberry.mutation
+    @strawberry.mutation(permission_classes=[IsNotReadOnly])  # type: ignore
     @classmethod
     async def chat_completion(
         cls, info: Info[Context, None], input: ChatCompletionInput
@@ -266,11 +282,19 @@ class ChatCompletionMutationMixin:
         provider_key = input.model.provider_key
         llm_client_class = PLAYGROUND_CLIENT_REGISTRY.get_client(provider_key, input.model.name)
         if llm_client_class is None:
-            raise BadRequest(f"No LLM client registered for provider '{provider_key}'")
-        llm_client = llm_client_class(
-            model=input.model,
-            api_key=input.api_key,
-        )
+            raise BadRequest(f"Unknown LLM provider: '{provider_key.value}'")
+        try:
+            llm_client = llm_client_class(
+                model=input.model,
+                api_key=input.api_key,
+            )
+        except CustomGraphQLError:
+            raise
+        except Exception as error:
+            raise BadRequest(
+                f"Failed to connect to LLM API for {provider_key.value} {input.model.name}: "
+                f"{str(error)}"
+            )
         return await cls._chat_completion(info, llm_client, input)
 
     @classmethod
@@ -459,6 +483,8 @@ def _template_formatter(template_language: TemplateLanguage) -> TemplateFormatte
         return MustacheTemplateFormatter()
     if template_language is TemplateLanguage.F_STRING:
         return FStringTemplateFormatter()
+    if template_language is TemplateLanguage.NONE:
+        return NoOpFormatter()
     assert_never(template_language)
 
 
@@ -486,6 +512,11 @@ def _llm_output_messages(
     if text_content:
         yield f"{LLM_OUTPUT_MESSAGES}.0.{MESSAGE_CONTENT}", text_content
     for tool_call_index, tool_call in enumerate(tool_calls.values()):
+        if tool_call_id := tool_call.id:
+            yield (
+                f"{LLM_OUTPUT_MESSAGES}.0.{MESSAGE_TOOL_CALLS}.{tool_call_index}.{TOOL_CALL_ID}",
+                tool_call_id,
+            )
         yield (
             f"{LLM_OUTPUT_MESSAGES}.0.{MESSAGE_TOOL_CALLS}.{tool_call_index}.{TOOL_CALL_FUNCTION_NAME}",
             tool_call.function.name,
@@ -513,6 +544,19 @@ def _serialize_event(event: SpanException) -> dict[str, Any]:
     return {k: (v.isoformat() if isinstance(v, datetime) else v) for k, v in asdict(event).items()}
 
 
+_AnyT = TypeVar("_AnyT")
+
+
+def _get_batches(
+    iterable: Iterable[_AnyT],
+    batch_size: int,
+) -> Iterator[list[_AnyT]]:
+    """Splits an iterable into batches not exceeding a specified size."""
+    iterator = iter(iterable)
+    while batch := list(islice(iterator, batch_size)):
+        yield batch
+
+
 JSON = OpenInferenceMimeTypeValues.JSON.value
 TEXT = OpenInferenceMimeTypeValues.TEXT.value
 LLM = OpenInferenceSpanKindValues.LLM.value
@@ -534,6 +578,7 @@ MESSAGE_CONTENT = MessageAttributes.MESSAGE_CONTENT
 MESSAGE_ROLE = MessageAttributes.MESSAGE_ROLE
 MESSAGE_TOOL_CALLS = MessageAttributes.MESSAGE_TOOL_CALLS
 
+TOOL_CALL_ID = ToolCallAttributes.TOOL_CALL_ID
 TOOL_CALL_FUNCTION_NAME = ToolCallAttributes.TOOL_CALL_FUNCTION_NAME
 TOOL_CALL_FUNCTION_ARGUMENTS_JSON = ToolCallAttributes.TOOL_CALL_FUNCTION_ARGUMENTS_JSON