-
-
Notifications
You must be signed in to change notification settings - Fork 95
Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
feat: Add ToolsAgent for models with tool-calling support (#530)
Co-authored-by: David Miguel <[email protected]>
- Loading branch information
1 parent
42c8d48
commit f3ee5b4
Showing
11 changed files
with
1,628 additions
and
11 deletions.
There are no files selected for viewing
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,190 @@ | ||
| # Tools Agent | ||
|
|
||
| An agent powered by the [tool calling API](/modules/model_io/models/chat_models/how_to/tools.md). | ||
|
|
||
| This agent is designed to work with any chat model that supports tool calling. It can interpret the model's output and decide when to call specific tools based on that output. | ||
|
|
||
| **Supported models:** | ||
| You can use any chat model that supports tool calling, like `ChatOpenAI`, `ChatOllama`, `ChatAnthropic`, `ChatFirebaseVertexAI`, etc. Check the [tool calling docs](/modules/model_io/models/chat_models/how_to/tools.md) for a complete list. | ||
|
|
||
| ## Usage | ||
|
|
||
| In the following example, we use `ChatOllama` with the `llama3.1` model and a calculator tool (included in `langchain_community`) to calculate the result of a mathematical expression. | ||
|
|
||
| ```dart | ||
| import 'package:langchain/langchain.dart'; | ||
| import 'package:langchain_community/langchain_community.dart'; | ||
| import 'package:langchain_ollama/langchain_ollama.dart'; | ||
| //... | ||
| final llm = ChatOllama( | ||
| defaultOptions: ChatOllamaOptions( | ||
| model: 'llama3.1', | ||
| temperature: 0, | ||
| ), | ||
| ); | ||
| final tool = CalculatorTool(); | ||
| final agent = ToolsAgent.fromLLMAndTools(llm: llm, tools: [tool]); | ||
| final executor = AgentExecutor(agent: agent); | ||
| final res = await executor.run( | ||
| 'What is 40 raised to the power of 0.43? ' | ||
| 'Return the result with 3 decimals.', | ||
| ); | ||
| print(res); | ||
| // The result is: 4.885 | ||
| ``` | ||
|
|
||
| ## Custom tools | ||
|
|
||
| You can easily call your own functions by wrapping them in a `Tool`. You can also add memory to the agent by passing it when creating the agent. | ||
|
|
||
| Let's see an example of how to do this. | ||
|
|
||
| First, let's create a class that will be the input for our tool. | ||
|
|
||
| ```dart | ||
| @immutable | ||
| class SearchInput { | ||
| const SearchInput({ | ||
| required this.query, | ||
| required this.n, | ||
| }); | ||
| final String query; | ||
| final int n; | ||
| SearchInput.fromJson(final Map<String, dynamic> json) | ||
| : this( | ||
| query: json['query'] as String, | ||
| n: json['n'] as int, | ||
| ); | ||
| } | ||
| ``` | ||
|
|
||
| Now let's define the tool: | ||
|
|
||
| ```dart | ||
| final searchTool = Tool.fromFunction<SearchInput, String>( | ||
| name: 'search', | ||
| description: 'Tool for searching the web.', | ||
| inputJsonSchema: { | ||
| 'type': 'object', | ||
| 'properties': { | ||
| 'query': { | ||
| 'type': 'string', | ||
| 'description': 'The query to search for', | ||
| }, | ||
| 'n': { | ||
| 'type': 'integer', | ||
| 'description': 'The number of results to return', | ||
| }, | ||
| }, | ||
| 'required': ['query'], | ||
| }, | ||
| func: callYourSearchFunction, | ||
| getInputFromJson: SearchInput.fromJson, | ||
| ); | ||
| ``` | ||
|
|
||
| Notice that we need to provide a function that converts the JSON input that the model will send to our tool into the input class that we defined. | ||
|
|
||
| The tool will call `callYourSearchFunction` function with the parsed input. For simplicity, we will just mock the search function. | ||
| ```dart | ||
| String callYourSearchFunction(final SearchInput input) { | ||
| final n = input.n; | ||
| final res = List<String>.generate( | ||
| n, | ||
| (i) => 'Result ${i + 1}: ${String.fromCharCode(65 + i) * 3}', | ||
| ); | ||
| return 'Results:\n${res.join('\n')}'; | ||
| } | ||
| ``` | ||
|
|
||
| Now we can create the agent and run it: | ||
|
|
||
| ```dart | ||
| final llm = ChatOllama( | ||
| defaultOptions: ChatOllamaOptions( | ||
| model: 'llama3-groq-tool-use', | ||
| temperature: 0, | ||
| ), | ||
| ); | ||
| final memory = ConversationBufferMemory(returnMessages: true); | ||
| final agent = ToolsAgent.fromLLMAndTools( | ||
| llm: llm, | ||
| tools: [searchTool], | ||
| memory: memory, | ||
| ); | ||
| final executor = AgentExecutor(agent: agent); | ||
| final res1 = await executor.run( | ||
| 'Search for cat names. Return only 3 results.', | ||
| ); | ||
| print(res1); | ||
| // Here are the top 3 cat names I found: AAA, BBB, and CCC. | ||
| ``` | ||
|
|
||
| ## Custom agent using LangChain Expression Language (LCEL) | ||
|
|
||
| You can replicate the functionality of the Tools Agent by using the LangChain Expression Language (LCEL) directly. | ||
|
|
||
| ```dart | ||
| final openAiKey = Platform.environment['OPENAI_API_KEY']; | ||
| final prompt = ChatPromptTemplate.fromTemplates([ | ||
| (ChatMessageType.system, 'You are a helpful assistant'), | ||
| (ChatMessageType.human, '{input}'), | ||
| (ChatMessageType.messagesPlaceholder, 'agent_scratchpad'), | ||
| ]); | ||
| final tool = CalculatorTool(); | ||
| final model = ChatOpenAI( | ||
| apiKey: openAiKey, | ||
| defaultOptions: ChatOpenAIOptions( | ||
| model: 'gpt-4o-mini', | ||
| temperature: 0, | ||
| tools: [tool], | ||
| ), | ||
| ); | ||
| const outputParser = ToolsAgentOutputParser(); | ||
| List<ChatMessage> buildScratchpad(final List<AgentStep> intermediateSteps) { | ||
| return intermediateSteps | ||
| .map((s) { | ||
| return s.action.messageLog + | ||
| [ | ||
| ChatMessage.tool( | ||
| toolCallId: s.action.id, | ||
| content: s.observation, | ||
| ), | ||
| ]; | ||
| }) | ||
| .expand((m) => m) | ||
| .toList(growable: false); | ||
| } | ||
| final agent = Agent.fromRunnable( | ||
| Runnable.mapInput( | ||
| (AgentPlanInput planInput) => <String, dynamic>{ | ||
| 'input': planInput.inputs['input'], | ||
| 'agent_scratchpad': buildScratchpad(planInput.intermediateSteps), | ||
| }, | ||
| ).pipe(prompt).pipe(model).pipe(outputParser), | ||
| tools: [tool], | ||
| ); | ||
| final executor = AgentExecutor(agent: agent); | ||
| final res = await executor.invoke({ | ||
| 'input': 'What is 40 raised to the power of 0.43? ' | ||
| 'Return the result with 3 decimals.', | ||
| }); | ||
| print(res['output']); | ||
| // The result of 40 raised to the power of 0.43 is approximately 4.885. | ||
| ``` | ||
|
|
||
| In this way, you can create your own custom agents with full control over their behavior, while still leveraging the flexibility of the Tools Agent to work with various language models and tools. |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Oops, something went wrong.