partners: AI21 Labs Jamba Streaming Support (#23538)

Thank you for contributing to LangChain!

- [x] **PR title**: "package: description"

- [x] **PR message**: ***Delete this entire checklist*** and replace
with
    - **Description:** Added support for streaming in AI21 Jamba Model
    - **Twitter handle:** https://github.com/AI21Labs


- [x] **Add tests and docs**: If you're adding a new integration, please
include

- [x] **Lint and test**: Run `make format`, `make lint` and `make test`
from the root of the package(s) you've modified. See contribution
guidelines for more: https://python.langchain.com/docs/contributing/

Additional guidelines:
- Make sure optional dependencies are imported within a function.
- Please do not add dependencies to pyproject.toml files (even optional
ones) unless they are required for unit tests.
- Most PRs should not touch more than one package.
- Changes should be backwards compatible.
- If you are adding something to community, do not re-import it in
langchain.

---------

Co-authored-by: Asaf Gardin <[email protected]>
Co-authored-by: Erick Friis <[email protected]>
Co-authored-by: Chester Curme <[email protected]>

libs/partners/ai21/README.md

@@ -28,13 +28,40 @@ Then initialize
 from langchain_core.messages import HumanMessage
 from langchain_ai21.chat_models import ChatAI21
 
-chat = ChatAI21(model="jamba-instruct-preview")
+chat = ChatAI21(model="jamba-instruct")
 messages = [HumanMessage(content="Hello from AI21")]
 chat.invoke(messages)
 ```
 
 For a list of the supported models, see [this page](https://docs.ai21.com/reference/python-sdk#chat)
 
+### Streaming in Chat
+Streaming is supported by the latest models. To use streaming, set the `streaming` parameter to `True` when initializing the model.
+
+```python
+from langchain_core.messages import HumanMessage
+from langchain_ai21.chat_models import ChatAI21
+
+chat = ChatAI21(model="jamba-instruct", streaming=True)
+messages = [HumanMessage(content="Hello from AI21")]
+
+response = chat.invoke(messages)
+```
+
+or use the `stream` method directly
+
+```python
+from langchain_core.messages import HumanMessage
+from langchain_ai21.chat_models import ChatAI21
+
+chat = ChatAI21(model="jamba-instruct")
+messages = [HumanMessage(content="Hello from AI21")]
+
+for chunk in chat.stream(messages):
+    print(chunk)
+```
+
+
 ## LLMs
 You can use AI21's Jurassic generative AI models as LangChain LLMs.
 To use the newer Jamba model, use the [ChatAI21 chat model](#chat-models), which

libs/partners/ai21/langchain_ai21/chat/chat_adapter.py

@@ -1,12 +1,21 @@
 from __future__ import annotations
 
 from abc import ABC, abstractmethod
-from typing import Any, Dict, List, Union, cast
+from typing import Any, Dict, Iterator, List, Literal, Union, cast, overload
 
 from ai21.models import ChatMessage as J2ChatMessage
 from ai21.models import RoleType
-from ai21.models.chat import ChatMessage
-from langchain_core.messages import AIMessage, BaseMessage, HumanMessage
+from ai21.models.chat import ChatCompletionChunk, ChatMessage
+from ai21.stream.stream import Stream as AI21Stream
+from langchain_core.messages import (
+    AIMessage,
+    AIMessageChunk,
+    BaseMessage,
+    BaseMessageChunk,
+    HumanMessage,
+)
+from langchain_core.messages.ai import UsageMetadata
+from langchain_core.outputs import ChatGenerationChunk
 
 _ChatMessageTypes = Union[ChatMessage, J2ChatMessage]
 _SYSTEM_ERR_MESSAGE = "System message must be at beginning of message list."
@@ -63,8 +72,31 @@ def _chat_message(
     ) -> _ChatMessageTypes:
         pass
 
+    @overload
+    def call(
+        self,
+        client: Any,
+        stream: Literal[True],
+        **params: Any,
+    ) -> Iterator[ChatGenerationChunk]:
+        pass
+
+    @overload
+    def call(
+        self,
+        client: Any,
+        stream: Literal[False],
+        **params: Any,
+    ) -> List[BaseMessage]:
+        pass
+
     @abstractmethod
-    def call(self, client: Any, **params: Any) -> List[BaseMessage]:
+    def call(
+        self,
+        client: Any,
+        stream: Literal[True] | Literal[False],
+        **params: Any,
+    ) -> List[BaseMessage] | Iterator[ChatGenerationChunk]:
         pass
 
     def _get_system_message_from_message(self, message: BaseMessage) -> str:
@@ -102,7 +134,33 @@ def _chat_message(
     ) -> J2ChatMessage:
         return J2ChatMessage(role=RoleType(role), text=content)
 
-    def call(self, client: Any, **params: Any) -> List[BaseMessage]:
+    @overload
+    def call(
+        self,
+        client: Any,
+        stream: Literal[True],
+        **params: Any,
+    ) -> Iterator[ChatGenerationChunk]:
+        ...
+
+    @overload
+    def call(
+        self,
+        client: Any,
+        stream: Literal[False],
+        **params: Any,
+    ) -> List[BaseMessage]:
+        ...
+
+    def call(
+        self,
+        client: Any,
+        stream: Literal[True] | Literal[False],
+        **params: Any,
+    ) -> List[BaseMessage] | Iterator[ChatGenerationChunk]:
+        if stream:
+            raise NotImplementedError("Streaming is not supported for Jurassic models.")
+
         response = client.chat.create(**params)
 
         return [AIMessage(output.text) for output in response.outputs]
@@ -128,7 +186,62 @@ def _chat_message(
             content=content,
         )
 
-    def call(self, client: Any, **params: Any) -> List[BaseMessage]:
-        response = client.chat.completions.create(**params)
+    @overload
+    def call(
+        self,
+        client: Any,
+        stream: Literal[True],
+        **params: Any,
+    ) -> Iterator[ChatGenerationChunk]:
+        ...
+
+    @overload
+    def call(
+        self,
+        client: Any,
+        stream: Literal[False],
+        **params: Any,
+    ) -> List[BaseMessage]:
+        ...
+
+    def call(
+        self,
+        client: Any,
+        stream: Literal[True] | Literal[False],
+        **params: Any,
+    ) -> List[BaseMessage] | Iterator[ChatGenerationChunk]:
+        response = client.chat.completions.create(stream=stream, **params)
+
+        if stream:
+            return self._stream_response(response)
 
         return [AIMessage(choice.message.content) for choice in response.choices]
+
+    def _stream_response(
+        self,
+        response: AI21Stream[ChatCompletionChunk],
+    ) -> Iterator[ChatGenerationChunk]:
+        for chunk in response:
+            converted_message = self._convert_ai21_chunk_to_chunk(chunk)
+            yield ChatGenerationChunk(message=converted_message)
+
+    def _convert_ai21_chunk_to_chunk(
+        self,
+        chunk: ChatCompletionChunk,
+    ) -> BaseMessageChunk:
+        usage = chunk.usage
+        content = chunk.choices[0].delta.content or ""
+
+        if usage is None:
+            return AIMessageChunk(
+                content=content,
+            )
+
+        return AIMessageChunk(
+            content=content,
+            usage_metadata=UsageMetadata(
+                input_tokens=usage.prompt_tokens,
+                output_tokens=usage.completion_tokens,
+                total_tokens=usage.total_tokens,
+            ),
+        )

libs/partners/ai21/langchain_ai21/chat_models.py

@@ -1,16 +1,20 @@
 import asyncio
 from functools import partial
-from typing import Any, Dict, List, Mapping, Optional
+from typing import Any, Dict, Iterator, List, Mapping, Optional
 
 from langchain_core.callbacks import (
     AsyncCallbackManagerForLLMRun,
     CallbackManagerForLLMRun,
 )
-from langchain_core.language_models.chat_models import BaseChatModel, LangSmithParams
+from langchain_core.language_models.chat_models import (
+    BaseChatModel,
+    LangSmithParams,
+    generate_from_stream,
+)
 from langchain_core.messages import (
     BaseMessage,
 )
-from langchain_core.outputs import ChatGeneration, ChatResult
+from langchain_core.outputs import ChatGeneration, ChatGenerationChunk, ChatResult
 from langchain_core.pydantic_v1 import root_validator
 
 from langchain_ai21.ai21_base import AI21Base
@@ -75,6 +79,7 @@ class ChatAI21(BaseChatModel, AI21Base):
 
     n: int = 1
     """Number of chat completions to generate for each prompt."""
+    streaming: bool = False
 
     _chat_adapter: ChatAdapter
 
@@ -166,14 +171,65 @@ def _generate(
         messages: List[BaseMessage],
         stop: Optional[List[str]] = None,
         run_manager: Optional[CallbackManagerForLLMRun] = None,
+        stream: Optional[bool] = None,
         **kwargs: Any,
     ) -> ChatResult:
-        params = self._build_params_for_request(messages=messages, stop=stop, **kwargs)
+        should_stream = stream or self.streaming
+
+        if should_stream:
+            return self._handle_stream_from_generate(
+                messages=messages,
+                stop=stop,
+                run_manager=run_manager,
+                **kwargs,
+            )
+
+        params = self._build_params_for_request(
+            messages=messages,
+            stop=stop,
+            stream=should_stream,
+            **kwargs,
+        )
+
         messages = self._chat_adapter.call(self.client, **params)
         generations = [ChatGeneration(message=message) for message in messages]
 
         return ChatResult(generations=generations)
 
+    def _handle_stream_from_generate(
+        self,
+        messages: List[BaseMessage],
+        stop: Optional[List[str]] = None,
+        run_manager: Optional[CallbackManagerForLLMRun] = None,
+        **kwargs: Any,
+    ) -> ChatResult:
+        stream_iter = self._stream(
+            messages=messages,
+            stop=stop,
+            run_manager=run_manager,
+            **kwargs,
+        )
+        return generate_from_stream(stream_iter)
+
+    def _stream(
+        self,
+        messages: List[BaseMessage],
+        stop: Optional[List[str]] = None,
+        run_manager: Optional[CallbackManagerForLLMRun] = None,
+        **kwargs: Any,
+    ) -> Iterator[ChatGenerationChunk]:
+        params = self._build_params_for_request(
+            messages=messages,
+            stop=stop,
+            stream=True,
+            **kwargs,
+        )
+
+        for chunk in self._chat_adapter.call(self.client, **params):
+            if run_manager and isinstance(chunk.message.content, str):
+                run_manager.on_llm_new_token(token=chunk.message.content, chunk=chunk)
+            yield chunk
+
     async def _agenerate(
         self,
         messages: List[BaseMessage],

libs/partners/ai21/pyproject.toml

@@ -14,7 +14,7 @@ license = "MIT"
 python = ">=3.8.1,<4.0"
 langchain-core = "^0.2.4"
 langchain-text-splitters = "^0.2.0"
-ai21 = "^2.4.1"
+ai21 = "^2.7.0"
 
 [tool.poetry.group.test]
 optional = true