Add Anthropic ChatModel to langchain (#2293)

* Adds an Anthropic ChatModel
* Factors out common code in our LLMModel and ChatModel
* Supports streaming llm-tokens to the callbacks on a delta basis (until
a future V2 API does that for us)
* Some fixes

langchain/chat_models/anthropic.py

@@ -0,0 +1,145 @@
+from typing import List, Optional
+
+from pydantic import Extra
+
+from langchain.chat_models.base import BaseChatModel
+from langchain.llms.anthropic import _AnthropicCommon
+from langchain.schema import (
+    AIMessage,
+    BaseMessage,
+    ChatGeneration,
+    ChatMessage,
+    ChatResult,
+    HumanMessage,
+    SystemMessage,
+)
+
+
+class ChatAnthropic(BaseChatModel, _AnthropicCommon):
+    r"""Wrapper around Anthropic's large language model.
+
+    To use, you should have the ``anthropic`` python package installed, and the
+    environment variable ``ANTHROPIC_API_KEY`` set with your API key, or pass
+    it as a named parameter to the constructor.
+
+    Example:
+        .. code-block:: python
+            import anthropic
+            from langchain.llms import Anthropic
+            model = Anthropic(model="<model_name>", anthropic_api_key="my-api-key")
+
+            # Simplest invocation, automatically wrapped with HUMAN_PROMPT
+            # and AI_PROMPT.
+            response = model("What are the biggest risks facing humanity?")
+
+            # Or if you want to use the chat mode, build a few-shot-prompt, or
+            # put words in the Assistant's mouth, use HUMAN_PROMPT and AI_PROMPT:
+            raw_prompt = "What are the biggest risks facing humanity?"
+            prompt = f"{anthropic.HUMAN_PROMPT} {prompt}{anthropic.AI_PROMPT}"
+            response = model(prompt)
+    """
+
+    class Config:
+        """Configuration for this pydantic object."""
+
+        extra = Extra.forbid
+
+    @property
+    def _llm_type(self) -> str:
+        """Return type of chat model."""
+        return "anthropic-chat"
+
+    def _convert_one_message_to_text(self, message: BaseMessage) -> str:
+        if isinstance(message, ChatMessage):
+            message_text = f"\n\n{message.role.capitalize()}: {message.content}"
+        elif isinstance(message, HumanMessage):
+            message_text = f"{self.HUMAN_PROMPT} {message.content}"
+        elif isinstance(message, AIMessage):
+            message_text = f"{self.AI_PROMPT} {message.content}"
+        elif isinstance(message, SystemMessage):
+            message_text = f"{self.HUMAN_PROMPT} <admin>{message.content}</admin>"
+        else:
+            raise ValueError(f"Got unknown type {message}")
+        return message_text
+
+    def _convert_messages_to_text(self, messages: List[BaseMessage]) -> str:
+        """Format a list of strings into a single string with necessary newlines.
+
+        Args:
+            messages (List[BaseMessage]): List of BaseMessage to combine.
+
+        Returns:
+            str: Combined string with necessary newlines.
+        """
+        return "".join(
+            self._convert_one_message_to_text(message) for message in messages
+        )
+
+    def _convert_messages_to_prompt(self, messages: List[BaseMessage]) -> str:
+        """Format a list of messages into a full prompt for the Anthropic model
+
+        Args:
+            messages (List[BaseMessage]): List of BaseMessage to combine.
+
+        Returns:
+            str: Combined string with necessary HUMAN_PROMPT and AI_PROMPT tags.
+        """
+        if not self.AI_PROMPT:
+            raise NameError("Please ensure the anthropic package is loaded")
+
+        if not isinstance(messages[-1], AIMessage):
+            messages.append(AIMessage(content=""))
+        text = self._convert_messages_to_text(messages)
+        return (
+            text.rstrip()
+        )  # trim off the trailing ' ' that might come from the "Assistant: "
+
+    def _generate(
+        self, messages: List[BaseMessage], stop: Optional[List[str]] = None
+    ) -> ChatResult:
+        prompt = self._convert_messages_to_prompt(messages)
+        params = {"prompt": prompt, "stop_sequences": stop, **self._default_params}
+
+        if self.streaming:
+            completion = ""
+            stream_resp = self.client.completion_stream(**params)
+            for data in stream_resp:
+                delta = data["completion"][len(completion) :]
+                completion = data["completion"]
+                self.callback_manager.on_llm_new_token(
+                    delta,
+                    verbose=self.verbose,
+                )
+        else:
+            response = self.client.completion(**params)
+            completion = response["completion"]
+        message = AIMessage(content=completion)
+        return ChatResult(generations=[ChatGeneration(message=message)])
+
+    async def _agenerate(
+        self, messages: List[BaseMessage], stop: Optional[List[str]] = None
+    ) -> ChatResult:
+        prompt = self._convert_messages_to_prompt(messages)
+        params = {"prompt": prompt, "stop_sequences": stop, **self._default_params}
+
+        if self.streaming:
+            completion = ""
+            stream_resp = await self.client.acompletion_stream(**params)
+            async for data in stream_resp:
+                delta = data["completion"][len(completion) :]
+                completion = data["completion"]
+                if self.callback_manager.is_async:
+                    await self.callback_manager.on_llm_new_token(
+                        delta,
+                        verbose=self.verbose,
+                    )
+                else:
+                    self.callback_manager.on_llm_new_token(
+                        delta,
+                        verbose=self.verbose,
+                    )
+        else:
+            response = await self.client.acompletion(**params)
+            completion = response["completion"]
+        message = AIMessage(content=completion)
+        return ChatResult(generations=[ChatGeneration(message=message)])

langchain/llms/anthropic.py

@@ -1,51 +1,28 @@
 """Wrapper around Anthropic APIs."""
 import re
-from typing import Any, Dict, Generator, List, Mapping, Optional
+from typing import Any, Callable, Dict, Generator, List, Mapping, Optional
 
-from pydantic import Extra, root_validator
+from pydantic import BaseModel, Extra, root_validator
 
 from langchain.llms.base import LLM
 from langchain.utils import get_from_dict_or_env
 
 
-class Anthropic(LLM):
+class _AnthropicCommon(BaseModel):
-    r"""Wrapper around Anthropic large language models.
+    client: Any = None  #: :meta private:
-
+    model: str = "claude-latest"
-    To use, you should have the ``anthropic`` python package installed, and the
-    environment variable ``ANTHROPIC_API_KEY`` set with your API key, or pass
-    it as a named parameter to the constructor.
-
-    Example:
-        .. code-block:: python
-            import anthropic
-            from langchain.llms import Anthropic
-            model = Anthropic(model="<model_name>", anthropic_api_key="my-api-key")
-
-            # Simplest invocation, automatically wrapped with HUMAN_PROMPT
-            # and AI_PROMPT.
-            response = model("What are the biggest risks facing humanity?")
-
-            # Or if you want to use the chat mode, build a few-shot-prompt, or
-            # put words in the Assistant's mouth, use HUMAN_PROMPT and AI_PROMPT:
-            raw_prompt = "What are the biggest risks facing humanity?"
-            prompt = f"{anthropic.HUMAN_PROMPT} {prompt}{anthropic.AI_PROMPT}"
-            response = model(prompt)
-    """
-
-    client: Any  #: :meta private:
-    model: str = "claude-v1"
     """Model name to use."""
 
     max_tokens_to_sample: int = 256
     """Denotes the number of tokens to predict per generation."""
 
-    temperature: float = 1.0
+    temperature: Optional[float] = None
     """A non-negative float that tunes the degree of randomness in generation."""
 
-    top_k: int = 0
+    top_k: Optional[int] = None
     """Number of most likely tokens to consider at each step."""
 
-    top_p: float = 1
+    top_p: Optional[float] = None
     """Total probability mass of tokens to consider at each step."""
 
     streaming: bool = False
@@ -55,11 +32,7 @@ class Anthropic(LLM):
 
     HUMAN_PROMPT: Optional[str] = None
     AI_PROMPT: Optional[str] = None
-
+    count_tokens: Optional[Callable[[str], int]] = None
-    class Config:
-        """Configuration for this pydantic object."""
-
-        extra = Extra.forbid
 
     @root_validator()
     def validate_environment(cls, values: Dict) -> Dict:
@@ -73,32 +46,86 @@ class Anthropic(LLM):
             values["client"] = anthropic.Client(anthropic_api_key)
             values["HUMAN_PROMPT"] = anthropic.HUMAN_PROMPT
             values["AI_PROMPT"] = anthropic.AI_PROMPT
+            values["count_tokens"] = anthropic.count_tokens
         except ImportError:
             raise ValueError(
                 "Could not import anthropic python package. "
-                "Please install it with `pip install anthropic`."
+                "Please it install it with `pip install anthropic`."
             )
         return values
 
     @property
     def _default_params(self) -> Mapping[str, Any]:
         """Get the default parameters for calling Anthropic API."""
-        return {
+        d = {
             "max_tokens_to_sample": self.max_tokens_to_sample,
-            "temperature": self.temperature,
+            "model": self.model,
-            "top_k": self.top_k,
-            "top_p": self.top_p,
         }
+        if self.temperature is not None:
+            d["temperature"] = self.temperature
+        if self.top_k is not None:
+            d["top_k"] = self.top_k
+        if self.top_p is not None:
+            d["top_p"] = self.top_p
+        return d
 
     @property
     def _identifying_params(self) -> Mapping[str, Any]:
         """Get the identifying parameters."""
-        return {**{"model": self.model}, **self._default_params}
+        return {**{}, **self._default_params}
+
+    def _get_anthropic_stop(self, stop: Optional[List[str]] = None) -> List[str]:
+        if not self.HUMAN_PROMPT or not self.AI_PROMPT:
+            raise NameError("Please ensure the anthropic package is loaded")
+
+        if stop is None:
+            stop = []
+
+        # Never want model to invent new turns of Human / Assistant dialog.
+        stop.extend([self.HUMAN_PROMPT])
+
+        return stop
+
+    def get_num_tokens(self, text: str) -> int:
+        """Calculate number of tokens."""
+        if not self.count_tokens:
+            raise NameError("Please ensure the anthropic package is loaded")
+        return self.count_tokens(text)
+
+
+class Anthropic(LLM, _AnthropicCommon):
+    r"""Wrapper around Anthropic's large language models.
+
+    To use, you should have the ``anthropic`` python package installed, and the
+    environment variable ``ANTHROPIC_API_KEY`` set with your API key, or pass
+    it as a named parameter to the constructor.
+
+    Example:
+        .. code-block:: python
+            import anthropic
+            from langchain.llms import Anthropic
+            model = Anthropic(model="<model_name>", anthropic_api_key="my-api-key")
+
+            # Simplest invocation, automatically wrapped with HUMAN_PROMPT
+            # and AI_PROMPT.
+            response = model("What are the biggest risks facing humanity?")
+
+            # Or if you want to use the chat mode, build a few-shot-prompt, or
+            # put words in the Assistant's mouth, use HUMAN_PROMPT and AI_PROMPT:
+            raw_prompt = "What are the biggest risks facing humanity?"
+            prompt = f"{anthropic.HUMAN_PROMPT} {prompt}{anthropic.AI_PROMPT}"
+            response = model(prompt)
+    """
+
+    class Config:
+        """Configuration for this pydantic object."""
+
+        extra = Extra.forbid
 
     @property
     def _llm_type(self) -> str:
         """Return type of llm."""
-        return "anthropic"
+        return "anthropic-llm"
 
     def _wrap_prompt(self, prompt: str) -> str:
         if not self.HUMAN_PROMPT or not self.AI_PROMPT:
@@ -115,18 +142,6 @@ class Anthropic(LLM):
         # As a last resort, wrap the prompt ourselves to emulate instruct-style.
         return f"{self.HUMAN_PROMPT} {prompt}{self.AI_PROMPT} Sure, here you go:\n"
 
-    def _get_anthropic_stop(self, stop: Optional[List[str]] = None) -> List[str]:
-        if not self.HUMAN_PROMPT or not self.AI_PROMPT:
-            raise NameError("Please ensure the anthropic package is loaded")
-
-        if stop is None:
-            stop = []
-
-        # Never want model to invent new turns of Human / Assistant dialog.
-        stop.extend([self.HUMAN_PROMPT, self.AI_PROMPT])
-
-        return stop
-
     def _call(self, prompt: str, stop: Optional[List[str]] = None) -> str:
         r"""Call out to Anthropic's completion endpoint.
 
@@ -148,10 +163,8 @@ class Anthropic(LLM):
         stop = self._get_anthropic_stop(stop)
         if self.streaming:
             stream_resp = self.client.completion_stream(
-                model=self.model,
                 prompt=self._wrap_prompt(prompt),
                 stop_sequences=stop,
-                stream=True,
                 **self._default_params,
             )
             current_completion = ""
@@ -163,7 +176,6 @@ class Anthropic(LLM):
                 )
             return current_completion
         response = self.client.completion(
-            model=self.model,
             prompt=self._wrap_prompt(prompt),
             stop_sequences=stop,
             **self._default_params,
@@ -175,10 +187,8 @@ class Anthropic(LLM):
         stop = self._get_anthropic_stop(stop)
         if self.streaming:
             stream_resp = await self.client.acompletion_stream(
-                model=self.model,
                 prompt=self._wrap_prompt(prompt),
                 stop_sequences=stop,
-                stream=True,
                 **self._default_params,
             )
             current_completion = ""
@@ -195,7 +205,6 @@ class Anthropic(LLM):
                     )
             return current_completion
         response = await self.client.acompletion(
-            model=self.model,
             prompt=self._wrap_prompt(prompt),
             stop_sequences=stop,
             **self._default_params,
@@ -227,7 +236,6 @@ class Anthropic(LLM):
         """
         stop = self._get_anthropic_stop(stop)
         return self.client.completion_stream(
-            model=self.model,
             prompt=self._wrap_prompt(prompt),
             stop_sequences=stop,
             **self._default_params,

tests/integration_tests/chat_models/test_anthropic.py

@@ -0,0 +1,81 @@
+"""Test Anthropic API wrapper."""
+from typing import List
+
+import pytest
+
+from langchain.callbacks.base import CallbackManager
+from langchain.chat_models.anthropic import ChatAnthropic
+from langchain.schema import (
+    AIMessage,
+    BaseMessage,
+    ChatGeneration,
+    HumanMessage,
+    LLMResult,
+)
+from tests.unit_tests.callbacks.fake_callback_handler import FakeCallbackHandler
+
+
+def test_anthropic_call() -> None:
+    """Test valid call to anthropic."""
+    chat = ChatAnthropic(model="bare-nano-0")
+    message = HumanMessage(content="Hello")
+    response = chat([message])
+    assert isinstance(response, AIMessage)
+    assert isinstance(response.content, str)
+
+
+def test_anthropic_streaming() -> None:
+    """Test streaming tokens from anthropic."""
+    chat = ChatAnthropic(model="bare-nano-0", streaming=True)
+    message = HumanMessage(content="Hello")
+    response = chat([message])
+    assert isinstance(response, AIMessage)
+    assert isinstance(response.content, str)
+
+
+def test_anthropic_streaming_callback() -> None:
+    """Test that streaming correctly invokes on_llm_new_token callback."""
+    callback_handler = FakeCallbackHandler()
+    callback_manager = CallbackManager([callback_handler])
+    chat = ChatAnthropic(
+        streaming=True,
+        callback_manager=callback_manager,
+        verbose=True,
+    )
+    message = HumanMessage(content="Write me a sentence with 100 words.")
+    chat([message])
+    assert callback_handler.llm_streams > 1
+
+
+@pytest.mark.asyncio
+async def test_anthropic_async_streaming_callback() -> None:
+    """Test that streaming correctly invokes on_llm_new_token callback."""
+    callback_handler = FakeCallbackHandler()
+    callback_manager = CallbackManager([callback_handler])
+    chat = ChatAnthropic(
+        streaming=True,
+        callback_manager=callback_manager,
+        verbose=True,
+    )
+    chat_messages: List[BaseMessage] = [
+        HumanMessage(content="How many toes do dogs have?")
+    ]
+    result: LLMResult = await chat.agenerate([chat_messages])
+    assert callback_handler.llm_streams > 1
+    assert isinstance(result, LLMResult)
+    for response in result.generations[0]:
+        assert isinstance(response, ChatGeneration)
+        assert isinstance(response.text, str)
+        assert response.text == response.message.content
+
+
+def test_formatting() -> None:
+    chat = ChatAnthropic()
+
+    chat_messages: List[BaseMessage] = [HumanMessage(content="Hello")]
+    result = chat._convert_messages_to_prompt(chat_messages)
+    assert result == "\n\nHuman: Hello\n\nAssistant:"
+
+    chat_messages = [HumanMessage(content="Hello"), AIMessage(content="Answer:")]
+    result = chat._convert_messages_to_prompt(chat_messages)
+    assert result == "\n\nHuman: Hello\n\nAssistant: Answer:"

tests/integration_tests/llms/test_anthropic.py

@@ -32,7 +32,6 @@ def test_anthropic_streaming_callback() -> None:
     callback_handler = FakeCallbackHandler()
     callback_manager = CallbackManager([callback_handler])
     llm = Anthropic(
-        model="claude-v1",
         streaming=True,
         callback_manager=callback_manager,
         verbose=True,
@@ -55,7 +54,6 @@ async def test_anthropic_async_streaming_callback() -> None:
     callback_handler = FakeCallbackHandler()
     callback_manager = CallbackManager([callback_handler])
     llm = Anthropic(
-        model="claude-v1",
         streaming=True,
         callback_manager=callback_manager,
         verbose=True,