Anthropic updated how they report token counts during streaming today.
See changes to `MessageDeltaUsage` in [this
commit](2da00f26c5 (diff-1a396eba0cd9cd8952dcdb58049d3b13f6b7768ead1411888d66e28211f7bfc5)).
It's clean and simple to grab these fields from the final
`message_delta` event. However, some of them are typed as Optional, and
language
[here](e42451ab3f/src/anthropic/lib/streaming/_messages.py (L462))
suggests they may not always be present. So here we take the required
field from the `message_delta` event as we were doing previously, and
ignore the rest.
- Support features from recent update:
https://www.anthropic.com/news/token-saving-updates (mostly adding
support for built-in tools in `bind_tools`
- Add documentation around prompt caching, token-efficient tool use, and
built-in tools.
- Support thinking blocks in core's `convert_to_openai_messages` (pass
through instead of error)
- Ignore thinking blocks in ChatOpenAI (instead of error)
- Support Anthropic-style image blocks in ChatOpenAI
---
Standard integration tests include a `supports_anthropic_inputs`
property which is currently enabled only for tests on `ChatAnthropic`.
This test enforces compatibility with message histories of the form:
```
- system message
- human message
- AI message with tool calls specified only through `tool_use` content blocks
- human message containing `tool_result` and an additional `text` block
```
It additionally checks support for Anthropic-style image inputs if
`supports_image_inputs` is enabled.
Here we change this test, such that if you enable
`supports_anthropic_inputs`:
- You support AI messages with text and `tool_use` content blocks
- You support Anthropic-style image inputs (if `supports_image_inputs`
is enabled)
- You support thinking content blocks.
That is, we add a test case for thinking content blocks, but we also
remove the requirement of handling tool results within HumanMessages
(motivated by existing agent abstractions, which should all return
ToolMessage). We move that requirement to a ChatAnthropic-specific test.
Structured output will currently always raise a BadRequestError when
Claude 3.7 Sonnet's `thinking` is enabled, because we rely on forced
tool use for structured output and this feature is not supported when
`thinking` is enabled.
Here we:
- Emit a warning if `with_structured_output` is called when `thinking`
is enabled.
- Raise `OutputParserException` if no tool calls are generated.
This is arguably preferable to raising an error in all cases.
```python
from langchain_anthropic import ChatAnthropic
from pydantic import BaseModel
class Person(BaseModel):
name: str
age: int
llm = ChatAnthropic(
model="claude-3-7-sonnet-latest",
max_tokens=5_000,
thinking={"type": "enabled", "budget_tokens": 2_000},
)
structured_llm = llm.with_structured_output(Person) # <-- this generates a warning
```
```python
structured_llm.invoke("Alice is 30.") # <-- works
```
```python
structured_llm.invoke("Hello!") # <-- raises OutputParserException
```
Took a "census" of models supported by init_chat_model-- of those that
return model names in response metadata, these were the only two that
had it keyed under `"model"` instead of `"model_name"`.
We currently return string (and therefore no content blocks / citations)
if the response is of the form
```
[
{"text": "a claim", "citations": [...]},
]
```
There are other cases where we do return citations as-is:
```
[
{"text": "a claim", "citations": [...]},
{"text": "some other text"},
{"text": "another claim", "citations": [...]},
]
```
Here we update to return content blocks including citations in the first
case as well.
Last week Anthropic released version 0.39.0 of its python sdk, which
enabled support for Python 3.13. This release deleted a legacy
`client.count_tokens` method, which we currently access during init of
the `Anthropic` LLM. Anthropic has replaced this functionality with the
[client.beta.messages.count_tokens()
API](https://github.com/anthropics/anthropic-sdk-python/pull/726).
To enable support for `anthropic >= 0.39.0` and Python 3.13, here we
drop support for the legacy token counting method, and add support for
the new method via `ChatAnthropic.get_num_tokens_from_messages`.
To fully support the token counting API, we update the signature of
`get_num_tokens_from_message` to accept tools everywhere.
---------
Co-authored-by: Bagatur <22008038+baskaryan@users.noreply.github.com>
`ChatAnthropic` can get `stop_reason` from the resulting `AIMessage` in
`invoke` and `ainvoke`, but not in `stream` and `astream`.
This is a different behavior from `ChatOpenAI`.
It is possible to get `stop_reason` from `stream` as well, since it is
needed to determine the next action after the LLM call. This would be
easier to handle in situations where only `stop_reason` is needed.
- Issue: NA
- Dependencies: NA
- Twitter handle: https://x.com/kiarina37
- Refactor standard test classes to make them easier to configure
- Update openai to support stop_sequences init param
- Update groq to support stop_sequences init param
- Update fireworks to support max_retries init param
- Update ChatModel.bind_tools to type tool_choice
- Update groq to handle tool_choice="any". **this may be controversial**
---------
Co-authored-by: Chester Curme <chester.curme@gmail.com>
Preserves string content chunks for non tool call requests for
convenience.
One thing - Anthropic events look like this:
```
RawContentBlockStartEvent(content_block=TextBlock(text='', type='text'), index=0, type='content_block_start')
RawContentBlockDeltaEvent(delta=TextDelta(text='<thinking>\nThe', type='text_delta'), index=0, type='content_block_delta')
RawContentBlockDeltaEvent(delta=TextDelta(text=' provide', type='text_delta'), index=0, type='content_block_delta')
...
RawContentBlockStartEvent(content_block=ToolUseBlock(id='toolu_01GJ6x2ddcMG3psDNNe4eDqb', input={}, name='get_weather', type='tool_use'), index=1, type='content_block_start')
RawContentBlockDeltaEvent(delta=InputJsonDelta(partial_json='', type='input_json_delta'), index=1, type='content_block_delta')
```
Note that `delta` has a `type` field. With this implementation, I'm
dropping it because `merge_list` behavior will concatenate strings.
We currently have `index` as a special field when merging lists, would
it be worth adding `type` too?
If so, what do we set as a context block chunk? `text` vs.
`text_delta`/`tool_use` vs `input_json_delta`?
CC @ccurme @efriis @baskaryan
- Refactor streaming to use raw events;
- Add `stream_usage` class attribute and kwarg to stream methods that,
if True, will include separate chunks in the stream containing usage
metadata.
There are two ways to implement streaming with anthropic's python sdk.
They have slight differences in how they surface usage metadata.
1. [Use helper
functions](https://github.com/anthropics/anthropic-sdk-python?tab=readme-ov-file#streaming-helpers).
This is what we are doing now.
```python
count = 1
with client.messages.stream(**params) as stream:
for text in stream.text_stream:
snapshot = stream.current_message_snapshot
print(f"{count}: {snapshot.usage} -- {text}")
count = count + 1
final_snapshot = stream.get_final_message()
print(f"{count}: {final_snapshot.usage}")
```
```
1: Usage(input_tokens=8, output_tokens=1) -- Hello
2: Usage(input_tokens=8, output_tokens=1) -- !
3: Usage(input_tokens=8, output_tokens=1) -- How
4: Usage(input_tokens=8, output_tokens=1) -- can
5: Usage(input_tokens=8, output_tokens=1) -- I
6: Usage(input_tokens=8, output_tokens=1) -- assist
7: Usage(input_tokens=8, output_tokens=1) -- you
8: Usage(input_tokens=8, output_tokens=1) -- today
9: Usage(input_tokens=8, output_tokens=1) -- ?
10: Usage(input_tokens=8, output_tokens=12)
```
To do this correctly, we need to emit a new chunk at the end of the
stream containing the usage metadata.
2. [Handle raw
events](https://github.com/anthropics/anthropic-sdk-python?tab=readme-ov-file#streaming-responses)
```python
stream = client.messages.create(**params, stream=True)
count = 1
for event in stream:
print(f"{count}: {event}")
count = count + 1
```
```
1: RawMessageStartEvent(message=Message(id='msg_01Vdyov2kADZTXqSKkfNJXcS', content=[], model='claude-3-haiku-20240307', role='assistant', stop_reason=None, stop_sequence=None, type='message', usage=Usage(input_tokens=8, output_tokens=1)), type='message_start')
2: RawContentBlockStartEvent(content_block=TextBlock(text='', type='text'), index=0, type='content_block_start')
3: RawContentBlockDeltaEvent(delta=TextDelta(text='Hello', type='text_delta'), index=0, type='content_block_delta')
4: RawContentBlockDeltaEvent(delta=TextDelta(text='!', type='text_delta'), index=0, type='content_block_delta')
5: RawContentBlockDeltaEvent(delta=TextDelta(text=' How', type='text_delta'), index=0, type='content_block_delta')
6: RawContentBlockDeltaEvent(delta=TextDelta(text=' can', type='text_delta'), index=0, type='content_block_delta')
7: RawContentBlockDeltaEvent(delta=TextDelta(text=' I', type='text_delta'), index=0, type='content_block_delta')
8: RawContentBlockDeltaEvent(delta=TextDelta(text=' assist', type='text_delta'), index=0, type='content_block_delta')
9: RawContentBlockDeltaEvent(delta=TextDelta(text=' you', type='text_delta'), index=0, type='content_block_delta')
10: RawContentBlockDeltaEvent(delta=TextDelta(text=' today', type='text_delta'), index=0, type='content_block_delta')
11: RawContentBlockDeltaEvent(delta=TextDelta(text='?', type='text_delta'), index=0, type='content_block_delta')
12: RawContentBlockStopEvent(index=0, type='content_block_stop')
13: RawMessageDeltaEvent(delta=Delta(stop_reason='end_turn', stop_sequence=None), type='message_delta', usage=MessageDeltaUsage(output_tokens=12))
14: RawMessageStopEvent(type='message_stop')
```
Here we implement the second option, in part because it should make
things easier when implementing streaming tool calls in the near future.
This would add two new chunks to the stream-- one at the beginning and
one at the end-- with blank content and containing usage metadata. We
add kwargs to the stream methods and a class attribute allowing for this
behavior to be toggled. I enabled it by default. If we merge this we can
add the same kwargs / attribute to OpenAI.
Usage:
```python
from langchain_anthropic import ChatAnthropic
model = ChatAnthropic(
model="claude-3-haiku-20240307",
temperature=0
)
full = None
for chunk in model.stream("hi"):
full = chunk if full is None else full + chunk
print(chunk)
print(f"\nFull: {full}")
```
```
content='' id='run-8a20843f-25c7-4025-ad72-9add395899e3' usage_metadata={'input_tokens': 8, 'output_tokens': 0, 'total_tokens': 8}
content='Hello' id='run-8a20843f-25c7-4025-ad72-9add395899e3'
content='!' id='run-8a20843f-25c7-4025-ad72-9add395899e3'
content=' How' id='run-8a20843f-25c7-4025-ad72-9add395899e3'
content=' can' id='run-8a20843f-25c7-4025-ad72-9add395899e3'
content=' I' id='run-8a20843f-25c7-4025-ad72-9add395899e3'
content=' assist' id='run-8a20843f-25c7-4025-ad72-9add395899e3'
content=' you' id='run-8a20843f-25c7-4025-ad72-9add395899e3'
content=' today' id='run-8a20843f-25c7-4025-ad72-9add395899e3'
content='?' id='run-8a20843f-25c7-4025-ad72-9add395899e3'
content='' id='run-8a20843f-25c7-4025-ad72-9add395899e3' usage_metadata={'input_tokens': 0, 'output_tokens': 12, 'total_tokens': 12}
Full: content='Hello! How can I assist you today?' id='run-8a20843f-25c7-4025-ad72-9add395899e3' usage_metadata={'input_tokens': 8, 'output_tokens': 12, 'total_tokens': 20}
```
core[minor], langchain[patch], openai[minor], anthropic[minor], fireworks[minor], groq[minor], mistralai[minor]
```python
class ToolCall(TypedDict):
name: str
args: Dict[str, Any]
id: Optional[str]
class InvalidToolCall(TypedDict):
name: Optional[str]
args: Optional[str]
id: Optional[str]
error: Optional[str]
class ToolCallChunk(TypedDict):
name: Optional[str]
args: Optional[str]
id: Optional[str]
index: Optional[int]
class AIMessage(BaseMessage):
...
tool_calls: List[ToolCall] = []
invalid_tool_calls: List[InvalidToolCall] = []
...
class AIMessageChunk(AIMessage, BaseMessageChunk):
...
tool_call_chunks: Optional[List[ToolCallChunk]] = None
...
```
Important considerations:
- Parsing logic occurs within different providers;
- ~Changing output type is a breaking change for anyone doing explicit
type checking;~
- ~Langsmith rendering will need to be updated:
https://github.com/langchain-ai/langchainplus/pull/3561~
- ~Langserve will need to be updated~
- Adding chunks:
- ~AIMessage + ToolCallsMessage = ToolCallsMessage if either has
non-null .tool_calls.~
- Tool call chunks are appended, merging when having equal values of
`index`.
- additional_kwargs accumulate the normal way.
- During streaming:
- ~Messages can change types (e.g., from AIMessageChunk to
AIToolCallsMessageChunk)~
- Output parsers parse additional_kwargs (during .invoke they read off
tool calls).
Packages outside of `partners/`:
- https://github.com/langchain-ai/langchain-cohere/pull/7
- https://github.com/langchain-ai/langchain-google/pull/123/files
---------
Co-authored-by: Chester Curme <chester.curme@gmail.com>
