add unified tool call block

llama-index-core/llama_index/core/base/llms/types.py

@@ -443,6 +443,17 @@ class ThinkingBlock(BaseModel):
     )
 
 
+class ToolCallBlock(BaseModel):
+    block_type: Literal["tool_call"] = "tool_call"
+    tool_call_id: Optional[str] = Field(
+        default=None, description="ID of the tool call, if provided"
+    )
+    tool_name: str = Field(description="Name of the called tool")
+    tool_kwargs: dict[str, Any] | object | str = Field(
+        default_factory=dict, description="Arguments provided to the tool, if available"
+    )
+
+
 ContentBlock = Annotated[
     Union[
         TextBlock,
@@ -454,6 +465,7 @@ class ThinkingBlock(BaseModel):
         CitableBlock,
         CitationBlock,
         ThinkingBlock,
+        ToolCallBlock,
     ],
     Field(discriminator="block_type"),
 ]

llama-index-core/tests/base/llms/test_types.py

@@ -18,6 +18,7 @@
     CachePoint,
     CacheControl,
     ThinkingBlock,
+    ToolCallBlock,
 )
 from llama_index.core.bridge.pydantic import BaseModel
 from llama_index.core.bridge.pydantic import ValidationError
@@ -473,3 +474,19 @@ def test_thinking_block():
     assert block.additional_information == {"total_thinking_tokens": 1000}
     assert block.content == "hello world"
     assert block.num_tokens == 100
+
+
+def test_tool_call_block():
+    default_block = ToolCallBlock(tool_name="hello_world")
+    assert default_block.block_type == "tool_call"
+    assert default_block.tool_call_id is None
+    assert default_block.tool_name == "hello_world"
+    assert default_block.tool_kwargs == {}
+    custom_block = ToolCallBlock(
+        tool_name="hello_world",
+        tool_call_id="1",
+        tool_kwargs={"test": 1},
+    )
+    assert custom_block.tool_call_id == "1"
+    assert custom_block.tool_name == "hello_world"
+    assert custom_block.tool_kwargs == {"test": 1}

llama-index-integrations/llms/llama-index-llms-anthropic/llama_index/llms/anthropic/base.py

@@ -1,4 +1,3 @@
-import json
 from typing import (
     TYPE_CHECKING,
     Any,
@@ -12,6 +11,7 @@
     Set,
     Tuple,
     Union,
+    cast,
 )
 
 from llama_index.core.base.llms.types import (
@@ -22,6 +22,7 @@
     LLMMetadata,
     MessageRole,
     ContentBlock,
+    ToolCallBlock,
 )
 from llama_index.core.base.llms.types import TextBlock as LITextBlock
 from llama_index.core.base.llms.types import CitationBlock as LICitationBlock
@@ -42,6 +43,7 @@
     force_single_tool_call,
     is_function_calling_model,
     messages_to_anthropic_messages,
+    _anthropic_tool_call_to_tool_call_block,
 )
 
 import anthropic
@@ -344,8 +346,7 @@ def _completion_response_from_chat_response(
 
     def _get_blocks_and_tool_calls_and_thinking(
         self, response: Any
-    ) -> Tuple[List[ContentBlock], List[Dict[str, Any]], List[Dict[str, Any]]]:
-        tool_calls = []
+    ) -> Tuple[List[ContentBlock], List[Dict[str, Any]]]:
         blocks: List[ContentBlock] = []
         citations: List[TextCitation] = []
         tracked_citations: Set[str] = set()
@@ -385,9 +386,15 @@ def _get_blocks_and_tool_calls_and_thinking(
                     )
                 )
             elif isinstance(content_block, ToolUseBlock):
-                tool_calls.append(content_block.model_dump())
+                blocks.append(
+                    ToolCallBlock(
+                        tool_call_id=content_block.id,
+                        tool_name=content_block.name,
+                        tool_kwargs=content_block.input,
+                    )
+                )
 
-        return blocks, tool_calls, [x.model_dump() for x in citations]
+        return blocks, [x.model_dump() for x in citations]
 
     @llm_chat_callback()
     def chat(
@@ -405,17 +412,12 @@ def chat(
             **all_kwargs,
         )
 
-        blocks, tool_calls, citations = self._get_blocks_and_tool_calls_and_thinking(
-            response
-        )
+        blocks, citations = self._get_blocks_and_tool_calls_and_thinking(response)
 
         return AnthropicChatResponse(
             message=ChatMessage(
                 role=MessageRole.ASSISTANT,
                 blocks=blocks,
-                additional_kwargs={
-                    "tool_calls": tool_calls,
-                },
             ),
             citations=citations,
             raw=dict(response),
@@ -526,7 +528,12 @@ def gen() -> Generator[AnthropicChatResponse, None, None]:
                     yield AnthropicChatResponse(
                         message=ChatMessage(
                             role=role,
-                            blocks=content,
+                            blocks=[
+                                *content,
+                                *_anthropic_tool_call_to_tool_call_block(
+                                    cur_tool_calls
+                                ),
+                            ],
                             additional_kwargs={
                                 "tool_calls": [
                                     t.model_dump() for t in tool_calls_to_send
@@ -577,17 +584,12 @@ async def achat(
             **all_kwargs,
         )
 
-        blocks, tool_calls, citations = self._get_blocks_and_tool_calls_and_thinking(
-            response
-        )
+        blocks, citations = self._get_blocks_and_tool_calls_and_thinking(response)
 
         return AnthropicChatResponse(
             message=ChatMessage(
                 role=MessageRole.ASSISTANT,
                 blocks=blocks,
-                additional_kwargs={
-                    "tool_calls": tool_calls,
-                },
             ),
             citations=citations,
             raw=dict(response),
@@ -698,11 +700,12 @@ async def gen() -> ChatResponseAsyncGen:
                     yield AnthropicChatResponse(
                         message=ChatMessage(
                             role=role,
-                            blocks=content,
-                            additional_kwargs={
-                                "tool_calls": [t.dict() for t in tool_calls_to_send],
-                                "thinking": thinking.model_dump() if thinking else None,
-                            },
+                            blocks=[
+                                *content,
+                                *_anthropic_tool_call_to_tool_call_block(
+                                    cur_tool_calls
+                                ),
+                            ],
                         ),
                         citations=cur_citations,
                         delta=content_delta,
@@ -811,7 +814,11 @@ def get_tool_calls_from_response(
         **kwargs: Any,
     ) -> List[ToolSelection]:
         """Predict and call the tool."""
-        tool_calls = response.message.additional_kwargs.get("tool_calls", [])
+        tool_calls = [
+            block
+            for block in response.message.blocks
+            if isinstance(block, ToolCallBlock)
+        ]
 
         if len(tool_calls) < 1:
             if error_on_no_tool_call:
@@ -823,25 +830,13 @@ def get_tool_calls_from_response(
 
         tool_selections = []
         for tool_call in tool_calls:
-            if (
-                "input" not in tool_call
-                or "id" not in tool_call
-                or "name" not in tool_call
-            ):
-                raise ValueError("Invalid tool call.")
-            if tool_call["type"] != "tool_use":
-                raise ValueError("Invalid tool type. Unsupported by Anthropic")
-            argument_dict = (
-                json.loads(tool_call["input"])
-                if isinstance(tool_call["input"], str)
-                else tool_call["input"]
-            )
+            argument_dict = tool_call.tool_kwargs
 
             tool_selections.append(
                 ToolSelection(
-                    tool_id=tool_call["id"],
-                    tool_name=tool_call["name"],
-                    tool_kwargs=argument_dict,
+                    tool_id=tool_call.tool_call_id or "",
+                    tool_name=tool_call.tool_name,
+                    tool_kwargs=cast(Dict[str, Any], argument_dict),
                 )
             )
 

llama-index-integrations/llms/llama-index-llms-anthropic/llama_index/llms/anthropic/utils.py

@@ -15,6 +15,7 @@
     CitableBlock,
     CitationBlock,
     ThinkingBlock,
+    ToolCallBlock,
     ContentBlock,
 )
 
@@ -26,6 +27,7 @@
     ImageBlockParam,
     CacheControlEphemeralParam,
     Base64PDFSourceParam,
+    ToolUseBlock,
 )
 from anthropic.types import ContentBlockParam as AnthropicContentBlock
 from anthropic.types.beta import (
@@ -193,6 +195,19 @@ def _to_anthropic_document_block(block: DocumentBlock) -> DocumentBlockParam:
     )
 
 
+def _anthropic_tool_call_to_tool_call_block(tool_calls: list[ToolUseBlock]):
+    blocks = []
+    for tool_call in tool_calls:
+        blocks.append(
+            ToolCallBlock(
+                tool_call_id=tool_call.id,
+                tool_kwargs=tool_call.input,
+                tool_name=tool_call.name,
+            )
+        )
+    return blocks
+
+
 def blocks_to_anthropic_blocks(
     blocks: Sequence[ContentBlock], kwargs: dict[str, Any]
 ) -> List[AnthropicContentBlock]:
@@ -270,24 +285,18 @@ def blocks_to_anthropic_blocks(
         elif isinstance(block, CitationBlock):
             # No need to pass these back to Anthropic
             continue
+        elif isinstance(block, ToolCallBlock):
+            anthropic_blocks.append(
+                ToolUseBlockParam(
+                    id=block.tool_call_id or "",
+                    input=block.tool_kwargs,
+                    name=block.tool_name,
+                    type="tool_use",
+                )
+            )
         else:
             raise ValueError(f"Unsupported block type: {type(block)}")
 
-    tool_calls = kwargs.get("tool_calls", [])
-    for tool_call in tool_calls:
-        assert "id" in tool_call
-        assert "input" in tool_call
-        assert "name" in tool_call
-
-        anthropic_blocks.append(
-            ToolUseBlockParam(
-                id=tool_call["id"],
-                input=tool_call["input"],
-                name=tool_call["name"],
-                type="tool_use",
-            )
-        )
-
     return anthropic_blocks
 
 
@@ -346,6 +355,12 @@ def messages_to_anthropic_messages(
 
 
 def force_single_tool_call(response: ChatResponse) -> None:
-    tool_calls = response.message.additional_kwargs.get("tool_calls", [])
+    tool_calls = [
+        block for block in response.message.blocks if isinstance(block, ToolCallBlock)
+    ]
     if len(tool_calls) > 1:
-        response.message.additional_kwargs["tool_calls"] = [tool_calls[0]]
+        response.message.blocks = [
+            block
+            for block in response.message.blocks
+            if not isinstance(block, ToolCallBlock)
+        ] + [tool_calls[0]]

llama-index-integrations/llms/llama-index-llms-anthropic/pyproject.toml

@@ -27,7 +27,7 @@ dev = [
 
 [project]
 name = "llama-index-llms-anthropic"
-version = "0.9.1"
+version = "0.10.0"
 description = "llama-index llms anthropic integration"
 authors = [{name = "Your Name", email = "[email protected]"}]
 requires-python = ">=3.9,<4.0"

llama-index-integrations/llms/llama-index-llms-anthropic/tests/test_llms_anthropic.py

@@ -17,7 +17,7 @@
     CachePoint,
     CacheControl,
 )
-from llama_index.core.base.llms.types import ThinkingBlock
+from llama_index.core.base.llms.types import ThinkingBlock, ToolCallBlock
 from llama_index.core.tools import FunctionTool
 from llama_index.llms.anthropic import Anthropic
 from llama_index.llms.anthropic.base import AnthropicChatResponse
@@ -253,8 +253,16 @@ def test_tool_required():
         tool_required=True,
     )
     assert isinstance(response, AnthropicChatResponse)
-    assert response.message.additional_kwargs["tool_calls"] is not None
-    assert len(response.message.additional_kwargs["tool_calls"]) > 0
+    assert (
+        len(
+            [
+                block
+                for block in response.message.blocks
+                if isinstance(block, ToolCallBlock)
+            ]
+        )
+        > 0
+    )
 
     # Test with tool_required=False
     response = llm.chat_with_tools(
@@ -264,7 +272,16 @@ def test_tool_required():
     )
     assert isinstance(response, AnthropicChatResponse)
     # Should not use tools for a simple greeting
-    assert not response.message.additional_kwargs.get("tool_calls")
+    assert (
+        len(
+            [
+                block
+                for block in response.message.blocks
+                if isinstance(block, ToolCallBlock)
+            ]
+        )
+        == 0
+    )
 
     # should not blow up with no tools (regression test)
     response = llm.chat_with_tools(
@@ -273,7 +290,16 @@ def test_tool_required():
         tool_required=False,
     )
     assert isinstance(response, AnthropicChatResponse)
-    assert not response.message.additional_kwargs.get("tool_calls")
+    assert (
+        len(
+            [
+                block
+                for block in response.message.blocks
+                if isinstance(block, ToolCallBlock)
+            ]
+        )
+        == 0
+    )
 
 
 @pytest.mark.skipif(