llama-index-llms-openai 0.3.18__tar.gz → 0.3.19__tar.gz

{llama_index_llms_openai-0.3.18 → llama_index_llms_openai-0.3.19}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: llama-index-llms-openai
-Version: 0.3.18
+Version: 0.3.19
 Summary: llama-index llms openai integration
 License: MIT
 Author: llama-index
@@ -11,7 +11,7 @@ Classifier: Programming Language :: Python :: 3.9
 Classifier: Programming Language :: Python :: 3.10
 Classifier: Programming Language :: Python :: 3.11
 Classifier: Programming Language :: Python :: 3.12
-Requires-Dist: llama-index-core (>=0.12.4,<0.13.0)
+Requires-Dist: llama-index-core (>=0.12.17,<0.13.0)
 Requires-Dist: openai (>=1.58.1,<2.0.0)
 Description-Content-Type: text/markdown
 

{llama_index_llms_openai-0.3.18 → llama_index_llms_openai-0.3.19}/llama_index/llms/openai/base.py

@@ -222,6 +222,14 @@ class OpenAI(FunctionCallingLLM):
         default=None,
         description="The effort to use for reasoning models.",
     )
+    modalities: Optional[List[str]] = Field(
+        default=None,
+        description="The output modalities to use for the model.",
+    )
+    audio_config: Optional[Dict[str, Any]] = Field(
+        default=None,
+        description="The audio configuration to use for the model.",
+    )
 
     _client: Optional[SyncOpenAI] = PrivateAttr()
     _aclient: Optional[AsyncOpenAI] = PrivateAttr()
@@ -254,6 +262,8 @@ class OpenAI(FunctionCallingLLM):
         output_parser: Optional[BaseOutputParser] = None,
         strict: bool = False,
         reasoning_effort: Optional[Literal["low", "medium", "high"]] = None,
+        modalities: Optional[List[str]] = None,
+        audio_config: Optional[Dict[str, Any]] = None,
         **kwargs: Any,
     ) -> None:
         additional_kwargs = additional_kwargs or {}
@@ -288,6 +298,8 @@ class OpenAI(FunctionCallingLLM):
             output_parser=output_parser,
             strict=strict,
             reasoning_effort=reasoning_effort,
+            modalities=modalities,
+            audio_config=audio_config,
             **kwargs,
         )
 
@@ -375,6 +387,11 @@ class OpenAI(FunctionCallingLLM):
     def complete(
         self, prompt: str, formatted: bool = False, **kwargs: Any
     ) -> CompletionResponse:
+        if self.modalities and "audio" in self.modalities:
+            raise ValueError(
+                "Audio is not supported for completion. Use chat/achat instead."
+            )
+
         if self._use_chat_completions(kwargs):
             complete_fn = chat_to_completion_decorator(self._chat)
         else:
@@ -434,6 +451,11 @@ class OpenAI(FunctionCallingLLM):
             # O1 models support reasoning_effort of low, medium, high
             all_kwargs["reasoning_effort"] = self.reasoning_effort
 
+        if self.modalities is not None:
+            all_kwargs["modalities"] = self.modalities
+        if self.audio_config is not None:
+            all_kwargs["audio"] = self.audio_config
+
         return all_kwargs
 
     @llm_retry_decorator
@@ -459,7 +481,9 @@ class OpenAI(FunctionCallingLLM):
                 )
 
         openai_message = response.choices[0].message
-        message = from_openai_message(openai_message)
+        message = from_openai_message(
+            openai_message, modalities=self.modalities or ["text"]
+        )
         openai_token_logprobs = response.choices[0].logprobs
         logprobs = None
         if openai_token_logprobs and openai_token_logprobs.content:
@@ -476,6 +500,9 @@ class OpenAI(FunctionCallingLLM):
     def _stream_chat(
         self, messages: Sequence[ChatMessage], **kwargs: Any
     ) -> ChatResponseGen:
+        if self.modalities and "audio" in self.modalities:
+            raise ValueError("Audio is not supported for chat streaming")
+
         client = self._get_client()
         message_dicts = to_openai_message_dicts(
             messages,
@@ -667,6 +694,11 @@ class OpenAI(FunctionCallingLLM):
     async def acomplete(
         self, prompt: str, formatted: bool = False, **kwargs: Any
     ) -> CompletionResponse:
+        if self.modalities and "audio" in self.modalities:
+            raise ValueError(
+                "Audio is not supported for completion. Use chat/achat instead."
+            )
+
         if self._use_chat_completions(kwargs):
             acomplete_fn = achat_to_completion_decorator(self._achat)
         else:
@@ -708,7 +740,9 @@ class OpenAI(FunctionCallingLLM):
                 )
 
         openai_message = response.choices[0].message
-        message = from_openai_message(openai_message)
+        message = from_openai_message(
+            openai_message, modalities=self.modalities or ["text"]
+        )
         openai_token_logprobs = response.choices[0].logprobs
         logprobs = None
         if openai_token_logprobs and openai_token_logprobs.content:
@@ -725,6 +759,9 @@ class OpenAI(FunctionCallingLLM):
     async def _astream_chat(
         self, messages: Sequence[ChatMessage], **kwargs: Any
     ) -> ChatResponseAsyncGen:
+        if self.modalities and "audio" in self.modalities:
+            raise ValueError("Audio is not supported for chat streaming")
+
         aclient = self._get_aclient()
         message_dicts = to_openai_message_dicts(
             messages,

{llama_index_llms_openai-0.3.18 → llama_index_llms_openai-0.3.19}/llama_index/llms/openai/utils.py

@@ -27,6 +27,7 @@ from llama_index.core.base.llms.types import (
     LogProb,
     MessageRole,
     TextBlock,
+    AudioBlock,
 )
 from llama_index.core.bridge.pydantic import BaseModel
 
@@ -68,6 +69,11 @@ GPT4_MODELS: Dict[str, int] = {
     "gpt-4-turbo-2024-04-09": 128000,
     "gpt-4-turbo": 128000,
     "gpt-4o": 128000,
+    "gpt-4o-audio-preview": 128000,
+    "gpt-4o-audio-preview-2024-12-17": 128000,
+    "gpt-4o-audio-preview-2024-10-01": 128000,
+    "gpt-4o-mini-audio-preview": 128000,
+    "gpt-4o-mini-audio-preview-2024-12-17": 128000,
     "gpt-4o-2024-05-13": 128000,
     "gpt-4o-2024-08-06": 128000,
     "gpt-4o-2024-11-20": 128000,
@@ -270,7 +276,16 @@ def to_openai_message_dict(
     """Convert a ChatMessage to an OpenAI message dict."""
     content = []
     content_txt = ""
+    reference_audio_id = None
     for block in message.blocks:
+        if message.role == MessageRole.ASSISTANT:
+            reference_audio_id = message.additional_kwargs.get(
+                "reference_audio_id", None
+            )
+            # if reference audio id is provided, we don't need to send the audio
+            if reference_audio_id:
+                continue
+
         if isinstance(block, TextBlock):
             content.append({"type": "text", "text": block.text})
             content_txt += block.text
@@ -291,6 +306,18 @@ def to_openai_message_dict(
                         },
                     }
                 )
+        elif isinstance(block, AudioBlock):
+            audio_bytes = block.resolve_audio(as_base64=True).read()
+            audio_str = audio_bytes.decode("utf-8")
+            content.append(
+                {
+                    "type": "input_audio",
+                    "input_audio": {
+                        "data": audio_str,
+                        "format": block.format,
+                    },
+                }
+            )
         else:
             msg = f"Unsupported content block type: {type(block).__name__}"
             raise ValueError(msg)
@@ -304,19 +331,27 @@ def to_openai_message_dict(
         else content_txt
     )
 
-    # NOTE: Despite what the openai docs say, if the role is ASSISTANT, SYSTEM
-    # or TOOL, 'content' cannot be a list and must be string instead.
-    # Furthermore, if all blocks are text blocks, we can use the content_txt
-    # as the content. This will avoid breaking openai-like APIs.
-    message_dict = {
-        "role": message.role.value,
-        "content": (
-            content_txt
-            if message.role.value in ("assistant", "tool", "system")
-            or all(isinstance(block, TextBlock) for block in message.blocks)
-            else content
-        ),
-    }
+    # If reference audio id is provided, we don't need to send the audio
+    # NOTE: this is only a thing for assistant messages
+    if reference_audio_id:
+        message_dict = {
+            "role": message.role.value,
+            "audio": {"id": reference_audio_id},
+        }
+    else:
+        # NOTE: Despite what the openai docs say, if the role is ASSISTANT, SYSTEM
+        # or TOOL, 'content' cannot be a list and must be string instead.
+        # Furthermore, if all blocks are text blocks, we can use the content_txt
+        # as the content. This will avoid breaking openai-like APIs.
+        message_dict = {
+            "role": message.role.value,
+            "content": (
+                content_txt
+                if message.role.value in ("assistant", "tool", "system")
+                or all(isinstance(block, TextBlock) for block in message.blocks)
+                else content
+            ),
+        }
 
     # TODO: O1 models do not support system prompts
     if model is not None and model in O1_MODELS:
@@ -353,20 +388,29 @@ def to_openai_message_dicts(
     ]
 
 
-def from_openai_message(openai_message: ChatCompletionMessage) -> ChatMessage:
+def from_openai_message(
+    openai_message: ChatCompletionMessage, modalities: List[str]
+) -> ChatMessage:
     """Convert openai message dict to generic message."""
     role = openai_message.role
     # NOTE: Azure OpenAI returns function calling messages without a content key
-    content = openai_message.content
-
-    # function_call = None  # deprecated in OpenAI v 1.1.0
+    if "text" in modalities and openai_message.content:
+        blocks = [TextBlock(text=openai_message.content or "")]
+    else:
+        blocks = []
 
     additional_kwargs: Dict[str, Any] = {}
     if openai_message.tool_calls:
         tool_calls: List[ChatCompletionMessageToolCall] = openai_message.tool_calls
         additional_kwargs.update(tool_calls=tool_calls)
 
-    return ChatMessage(role=role, content=content, additional_kwargs=additional_kwargs)
+    if openai_message.audio and "audio" in modalities:
+        reference_audio_id = openai_message.audio.id
+        audio_data = openai_message.audio.data
+        additional_kwargs["reference_audio_id"] = reference_audio_id
+        blocks.append(AudioBlock(audio=audio_data, format="mp3"))
+
+    return ChatMessage(role=role, blocks=blocks, additional_kwargs=additional_kwargs)
 
 
 def from_openai_token_logprob(
@@ -421,10 +465,10 @@ def from_openai_completion_logprobs(
 
 
 def from_openai_messages(
-    openai_messages: Sequence[ChatCompletionMessage],
+    openai_messages: Sequence[ChatCompletionMessage], modalities: List[str]
 ) -> List[ChatMessage]:
     """Convert openai message dicts to generic messages."""
-    return [from_openai_message(message) for message in openai_messages]
+    return [from_openai_message(message, modalities) for message in openai_messages]
 
 
 def from_openai_message_dict(message_dict: dict) -> ChatMessage:

{llama_index_llms_openai-0.3.18 → llama_index_llms_openai-0.3.19}/pyproject.toml

@@ -29,12 +29,12 @@ exclude = ["**/BUILD"]
 license = "MIT"
 name = "llama-index-llms-openai"
 readme = "README.md"
-version = "0.3.18"
+version = "0.3.19"
 
 [tool.poetry.dependencies]
 python = ">=3.9,<4.0"
 openai = "^1.58.1"
-llama-index-core = "^0.12.4"
+llama-index-core = "^0.12.17"
 
 [tool.poetry.group.dev.dependencies]
 ipython = "8.10.0"