Skip to content

The agent engineering platform.

Small bug fixes (#23353)

joshc-ai2116a293cc3ae9
diff --git a/docs/docs/integrations/chat/ai21.ipynb b/docs/docs/integrations/chat/ai21.ipynb
index fed515dd5a..4a0d0f86e3 100644
--- a/docs/docs/integrations/chat/ai21.ipynb
+++ b/docs/docs/integrations/chat/ai21.ipynb
@@ -18,7 +18,9 @@
"# ChatAI21\n",
"\n",
"This notebook covers how to get started with AI21 chat models.\n",
- "\n",
+ "Note that different chat models support different parameters. See the ",
+ "[AI21 documentation](https://docs.ai21.com/reference) to learn more about the parameters in your chosen model.\n",
+ "[See all AI21's LangChain components.](https://pypi.org/project/langchain-ai21/) \n",
"## Installation"
]
},
@@ -44,7 +46,8 @@
"source": [
"## Environment Setup\n",
"\n",
- "We'll need to get a [AI21 API key](https://docs.ai21.com/) and set the `AI21_API_KEY` environment variable:\n"
+ "We'll need to get an [AI21 API key](https://docs.ai21.com/) and set the ",
+ "`AI21_API_KEY` environment variable:\n"
]
},
{
diff --git a/docs/docs/integrations/llms/ai21.ipynb b/docs/docs/integrations/llms/ai21.ipynb
index 72fbe840fe..da2ec1293c 100644
--- a/docs/docs/integrations/llms/ai21.ipynb
+++ b/docs/docs/integrations/llms/ai21.ipynb
@@ -17,7 +17,9 @@
"source": [
"# AI21LLM\n",
"\n",
- "This example goes over how to use LangChain to interact with `AI21` models.\n",
+ "This example goes over how to use LangChain to interact with `AI21` Jurassic models. To use the Jamba model, use the [ChatAI21 object](https://python.langchain.com/v0.2/docs/integrations/chat/ai21/) instead.\n",
+ "\n",
+ "[See a full list of AI21 models and tools on LangChain.](https://pypi.org/project/langchain-ai21/)\n",
"\n",
"## Installation"
]
diff --git a/libs/partners/ai21/README.md b/libs/partners/ai21/README.md
index 98b8815e98..524854c7c7 100644
--- a/libs/partners/ai21/README.md
+++ b/libs/partners/ai21/README.md
@@ -1,6 +1,6 @@
# langchain-ai21
-This package contains the LangChain integrations for [AI21](https://docs.ai21.com/) through their [AI21](https://pypi.org/project/ai21/) SDK.
+This package contains the LangChain integrations for [AI21](https://docs.ai21.com/) models and tools.
## Installation and Setup
@@ -13,9 +13,10 @@ pip install langchain-ai21
## Chat Models
-This package contains the `ChatAI21` class, which is the recommended way to interface with AI21 Chat models.
+This package contains the `ChatAI21` class, which is the recommended way to interface with AI21 chat models, including Jamba-Instruct
+and any Jurassic chat models.
-To use, install the requirements, and configure your environment.
+To use, install the requirements and configure your environment.
```bash
export AI21_API_KEY=your-api-key
@@ -27,7 +28,7 @@ Then initialize
from langchain_core.messages import HumanMessage
from langchain_ai21.chat_models import ChatAI21
-chat = ChatAI21(model="jamab-instruct")
+chat = ChatAI21(model="jamba-instruct-preview")
messages = [HumanMessage(content="Hello from AI21")]
chat.invoke(messages)
```
@@ -35,10 +36,12 @@ chat.invoke(messages)
For a list of the supported models, see [this page](https://docs.ai21.com/reference/python-sdk#chat)
## LLMs
-You can use AI21's generative AI models as Langchain 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
+supports single-turn instruction/question answering capabilities.
```python
-from langchain.prompts import PromptTemplate
+from langchain_core.prompts import PromptTemplate
from langchain_ai21 import AI21LLM
llm = AI21LLM(model="j2-ultra")
@@ -56,7 +59,7 @@ print(chain.invoke({"question": question}))
## Embeddings
-You can use AI21's embeddings models as:
+You can use AI21's [embeddings model](https://docs.ai21.com/reference/embeddings-ref) as shown here:
### Query
@@ -76,12 +79,12 @@ embeddings = AI21Embeddings()
embeddings.embed_documents(["Hello! This is document 1", "And this is document 2!"])
```
-## Task Specific Models
+## Task-Specific Models
### Contextual Answers
-You can use AI21's contextual answers model to receives text or document, serving as a context,
-and a question and returns an answer based entirely on this context.
+You can use AI21's [contextual answers model](https://docs.ai21.com/reference/contextual-answers-ref) to parse
+given text and answer a question based entirely on the provided information.
This means that if the answer to your question is not in the document,
the model will indicate it (instead of providing a false answer)
@@ -91,7 +94,7 @@ from langchain_ai21 import AI21ContextualAnswers
tsm = AI21ContextualAnswers()
-response = tsm.invoke(input={"context": "Your context", "question": "Your question"})
+response = tsm.invoke(input={"context": "Lots of information here", "question": "Your question about the context"})
```
You can also use it with chains and output parsers and vector DBs:
```python
@@ -110,8 +113,8 @@ response = chain.invoke(
### Semantic Text Splitter
-You can use AI21's semantic text splitter to split a text into segments.
-Instead of merely using punctuation and newlines to divide the text, it identifies distinct topics that will work well together and will form a coherent piece of text.
+You can use AI21's semantic [text segmentation model](https://docs.ai21.com/reference/text-segmentation-ref) to split a text into segments by topic.
+Text is split at each point where the topic changes.
For a list for examples, see [this page](https://github.com/langchain-ai/langchain/blob/master/docs/docs/modules/data_connection/document_transformers/semantic_text_splitter.ipynb).
diff --git a/libs/partners/ai21/langchain_ai21/chat_models.py b/libs/partners/ai21/langchain_ai21/chat_models.py
index 3cdafaca4f..9150ab4d88 100644
--- a/libs/partners/ai21/langchain_ai21/chat_models.py
+++ b/libs/partners/ai21/langchain_ai21/chat_models.py
@@ -19,7 +19,10 @@ from langchain_ai21.chat.chat_factory import create_chat_adapter
class ChatAI21(BaseChatModel, AI21Base):
- """ChatAI21 chat model.
+ """ChatAI21 chat model. Different model types support different parameters and
+ different parameter values. Please read the [AI21 reference documentation]
+ (https://docs.ai21.com/reference) for your model to understand which parameters
+ are available.
Example:
.. code-block:: python
@@ -27,7 +30,10 @@ class ChatAI21(BaseChatModel, AI21Base):
from langchain_ai21 import ChatAI21
- model = ChatAI21()
+ model = ChatAI21(
+ # defaults to os.environ.get("AI21_API_KEY")
+ api_key="my_api_key"
+ )
"""
model: str
@@ -42,7 +48,8 @@ class ChatAI21(BaseChatModel, AI21Base):
"""The maximum number of tokens to generate for each response."""
min_tokens: int = 0
- """The minimum number of tokens to generate for each response."""
+ """The minimum number of tokens to generate for each response.
+ _Not supported for all models._"""
temperature: float = 0.7
"""A value controlling the "creativity" of the model's responses."""
@@ -51,17 +58,20 @@ class ChatAI21(BaseChatModel, AI21Base):
"""A value controlling the diversity of the model's responses."""
top_k_return: int = 0
- """The number of top-scoring tokens to consider for each generation step."""
+ """The number of top-scoring tokens to consider for each generation step.
+ _Not supported for all models._"""
frequency_penalty: Optional[Any] = None
- """A penalty applied to tokens that are frequently generated."""
+ """A penalty applied to tokens that are frequently generated.
+ _Not supported for all models._"""
presence_penalty: Optional[Any] = None
- """ A penalty applied to tokens that are already present in the prompt."""
+ """ A penalty applied to tokens that are already present in the prompt.
+ _Not supported for all models._"""
count_penalty: Optional[Any] = None
"""A penalty applied to tokens based on their frequency
- in the generated responses."""
+ in the generated responses. _Not supported for all models._"""
n: int = 1
"""Number of chat completions to generate for each prompt."""
diff --git a/libs/partners/ai21/langchain_ai21/llms.py b/libs/partners/ai21/langchain_ai21/llms.py
index 0c0ae1822e..d1ca42e3dd 100644
--- a/libs/partners/ai21/langchain_ai21/llms.py
+++ b/libs/partners/ai21/langchain_ai21/llms.py
@@ -19,14 +19,24 @@ from langchain_ai21.ai21_base import AI21Base
class AI21LLM(BaseLLM, AI21Base):
- """AI21 large language models.
+ """AI21 large language models. Different model types support different parameters
+ and different parameter values. Please read the [AI21 reference documentation]
+ (https://docs.ai21.com/reference) for your model to understand which parameters
+ are available.
+
+ AI21LLM supports only the older Jurassic models.
+ We recommend using ChatAI21 with the newest models, for better results and more
+ features.
Example:
.. code-block:: python
from langchain_ai21 import AI21LLM
- model = AI21LLM()
+ model = AI21LLM(
+ # defaults to os.environ.get("AI21_API_KEY")
+ api_key="my_api_key"
+ )
"""
model: str
@@ -40,7 +50,8 @@ class AI21LLM(BaseLLM, AI21Base):
"""The maximum number of tokens to generate for each response."""
min_tokens: int = 0
- """The minimum number of tokens to generate for each response."""
+ """The minimum number of tokens to generate for each response.
+ _Not supported for all models._"""
temperature: float = 0.7
"""A value controlling the "creativity" of the model's responses."""
@@ -49,17 +60,20 @@ class AI21LLM(BaseLLM, AI21Base):
"""A value controlling the diversity of the model's responses."""
top_k_return: int = 0
- """The number of top-scoring tokens to consider for each generation step."""
+ """The number of top-scoring tokens to consider for each generation step.
+ _Not supported for all models._"""
frequency_penalty: Optional[Any] = None
- """A penalty applied to tokens that are frequently generated."""
+ """A penalty applied to tokens that are frequently generated.
+ _Not supported for all models._"""
presence_penalty: Optional[Any] = None
- """ A penalty applied to tokens that are already present in the prompt."""
+ """ A penalty applied to tokens that are already present in the prompt.
+ _Not supported for all models._"""
count_penalty: Optional[Any] = None
"""A penalty applied to tokens based on their frequency
- in the generated responses."""
+ in the generated responses. _Not supported for all models._"""
custom_model: Optional[str] = None
epoch: Optional[int] = None