github/langchain
1
0
Fork 0
mirror of https://github.com/hwchase17/langchain.git synced 2026-06-09 10:17:00 +00:00
Code Issues Packages Projects Releases Wiki Activity

ollama: thinking, tool streaming, docs, tests (#31772)

Browse Source 
* New `reasoning` (bool) param to support toggling [Ollama
thinking](https://ollama.com/blog/thinking) (#31573, #31700). If
`reasoning=True`, Ollama's `thinking` content will be placed in the
model responses' `additional_kwargs.reasoning_content`.
  * Supported by:
    * ChatOllama (class level, invocation level TODO)
    * OllamaLLM (TODO)
* Added tests to ensure streaming tool calls is successful (#29129)
* Refactored tests that relied on `extract_reasoning()`
* Myriad docs additions and consistency/typo fixes
* Improved type safety in some spots

Closes #29129
Addresses #31573 and #31700
Supersedes #31701
This commit is contained in:
Mason Daugherty
2025-07-07 13:56:41 -04:00
committed by GitHub
parent 0eb10f31c1
commit e686a70ee0
14 changed files with 630 additions and 213 deletions
Download Patch File Download Diff File Expand all files Collapse all files

67
libs/partners/ollama/langchain_ollama/llms.py
View File

@@ -36,6 +36,20 @@ class OllamaLLM(BaseLLM):
model: str
"""Model name to use."""
reasoning: Optional[bool] = True
"""Controls the reasoning/thinking mode for
`supported models <https://ollama.com/search?c=thinking>`__.
- ``True``: Enables reasoning mode. The model's reasoning process will be
captured and returned separately in the ``additional_kwargs`` of the
response message, under ``reasoning_content``. The main response
content will not include the reasoning tags.
- ``False``: Disables reasoning mode. The model will not perform any reasoning,
and the response will not include any reasoning content.
- ``None`` (Default): The model will use its default reasoning behavior. If
the model performs reasoning, the ``<think>`` and ``</think>`` tags will
be present directly within the main response content."""
validate_model_on_init: bool = False
"""Whether to validate the model exists in ollama locally on initialization."""
@@ -56,7 +70,7 @@ class OllamaLLM(BaseLLM):
num_ctx: Optional[int] = None
"""Sets the size of the context window used to generate the
next token. (Default: 2048) """
next token. (Default: 2048)"""
num_gpu: Optional[int] = None
"""The number of GPUs to use. On macOS it defaults to 1 to
@@ -137,12 +151,12 @@ class OllamaLLM(BaseLLM):
For a full list of the params, see the `HTTPX documentation <https://www.python-httpx.org/api/#client>`__.
"""
_client: Client = PrivateAttr(default=None) # type: ignore
_client: Optional[Client] = PrivateAttr(default=None)
"""
The client to use for making requests.
"""
_async_client: AsyncClient = PrivateAttr(default=None) # type: ignore
_async_client: Optional[AsyncClient] = PrivateAttr(default=None)
"""
The async client to use for making requests.
"""
@@ -155,7 +169,7 @@ class OllamaLLM(BaseLLM):
) -> dict[str, Any]:
if self.stop is not None and stop is not None:
raise ValueError("`stop` found in both the input and default params.")
elif self.stop is not None:
if self.stop is not None:
stop = self.stop
options_dict = kwargs.pop(
@@ -183,6 +197,7 @@ class OllamaLLM(BaseLLM):
"prompt": prompt,
"stream": kwargs.pop("stream", True),
"model": kwargs.pop("model", self.model),
"think": kwargs.pop("reasoning", self.reasoning),
"format": kwargs.pop("format", self.format),
"options": Options(**options_dict),
"keep_alive": kwargs.pop("keep_alive", self.keep_alive),
@@ -230,10 +245,11 @@ class OllamaLLM(BaseLLM):
stop: Optional[list[str]] = None,
**kwargs: Any,
) -> AsyncIterator[Union[Mapping[str, Any], str]]:
async for part in await self._async_client.generate(
**self._generate_params(prompt, stop=stop, **kwargs)
): # type: ignore
yield part # type: ignore
if self._async_client:
async for part in await self._async_client.generate(
**self._generate_params(prompt, stop=stop, **kwargs)
):
yield part
def _create_generate_stream(
self,
@@ -241,9 +257,10 @@ class OllamaLLM(BaseLLM):
stop: Optional[list[str]] = None,
**kwargs: Any,
) -> Iterator[Union[Mapping[str, Any], str]]:
yield from self._client.generate(
**self._generate_params(prompt, stop=stop, **kwargs)
) # type: ignore
if self._client:
yield from self._client.generate(
**self._generate_params(prompt, stop=stop, **kwargs)
)
async def _astream_with_aggregation(
self,
@@ -356,11 +373,19 @@ class OllamaLLM(BaseLLM):
) -> Iterator[GenerationChunk]:
for stream_resp in self._create_generate_stream(prompt, stop, **kwargs):
if not isinstance(stream_resp, str):
additional_kwargs = {}
if thinking_content := stream_resp.get("thinking"):
additional_kwargs["reasoning_content"] = thinking_content
chunk = GenerationChunk(
text=(stream_resp.get("response", "")),
generation_info=(
dict(stream_resp) if stream_resp.get("done") is True else None
),
generation_info={
"finish_reason": self.stop,
**additional_kwargs,
**(
dict(stream_resp) if stream_resp.get("done") is True else {}
),
},
)
if run_manager:
run_manager.on_llm_new_token(
@@ -378,11 +403,19 @@ class OllamaLLM(BaseLLM):
) -> AsyncIterator[GenerationChunk]:
async for stream_resp in self._acreate_generate_stream(prompt, stop, **kwargs):
if not isinstance(stream_resp, str):
additional_kwargs = {}
if thinking_content := stream_resp.get("thinking"):
additional_kwargs["reasoning_content"] = thinking_content
chunk = GenerationChunk(
text=(stream_resp.get("response", "")),
generation_info=(
dict(stream_resp) if stream_resp.get("done") is True else None
),
generation_info={
"finish_reason": self.stop,
**additional_kwargs,
**(
dict(stream_resp) if stream_resp.get("done") is True else {}
),
},
)
if run_manager:
await run_manager.on_llm_new_token(