Enterprises are increasingly adopting AI to enhance business workflows. AI models, equipped with external tools, have the potential to streamline business operations.

At Cohere, we’re excited to share that our latest generative models — Command R and R+ — have been trained to interact with a variety of tools. You can explore these tools capabilities via our API and via Langchain. Whether you want AI to automatically take action in your own enterprise systems, analyze internal documents and databases, conduct web searches, or write Python code, Cohere’s models can transform your business processes.

What Is Tool Use?

The term tool use refers to the ability of large language models to leverage external tools or APIs in order to handle a variety of tasks beyond text generation.

By chaining tools and API calls, AI models can reason, update enterprise systems, provide actionable insights, and automate business operations.

For example, by equipping Command R+ with your enterprise tools, it becomes possible to use natural language to:

Keep your CRM up-to-date by automatically analyzing sales call transcripts
Autonomously extract marketing insights from spreadsheets
Enable direct conversations with SQL databases to optimize procurement decisions
Power customer support bots that navigate company documentation, emails, and codebases to quickly resolve customer queries

Large language models with tool use are becoming the powerhouse to streamline business workflows.

How Does It Work?

Tool use comes in two flavors: single-step tool use and multi-step tool use.

Single-step tool use is useful for use cases where the model only needs to perform a single action or a set of independent actions that can happen simultaneously, such as private document search, sending an email, or setting up a calendar invite.

Multi-step tool use, on the other hand, is designed for situations where the model needs to carry out a sequence of actions, with each action building on the results of the previous ones. By enabling models to perform sequences of actions, we can automate and streamline a wide range of tasks, saving time and increasing efficiency.

For example, a financial analyst might want to understand possible investment strategies in a specific domain. The data is stored across internal databases, the internet, and private PDFs. Given a natural language request, the tool can perform a series of API calls to gather information from these various sources, analyze this data, and finally compile a succinct one-page analysis of the possible investment strategies.

Given a user request, the model comes up with a plan to solve the problem by answering questions, such as "Which tools should be used," and "In what order should they be used." Once the model decides it knows how to answer the user’s question, it proceeds to generate the final response. If the model cannot solve the problem after gathering as much information as it can, it will respond by saying so.

Tool use allows LLMs to go beyond text generation and into the world of AI agents — LLM-backed systems that can solve more difficult tasks by using tools over multiple steps.

Both single-step and multi-step tool use are designed to be user-friendly. To use these capabilities, you simply need to describe the tools available to the model, including what the tool does and its parameters. Once a user request is made, the model will create a plan and determine how to use these tools to fulfill the user request. After each tool has been called, the model has the ability to reason about the contents and can adapt the initial plan as necessary. The model for example can automatically retry calling an unknown API after it returns an error message.

This is an example of a tool description that the model can leverage to query daily sales data:

tools = [
   {
       "name": "query_daily_sales_report",
       "description": "Connects to a database to retrieve overall sales volumes and sales information for a given day.",
       "parameter_definitions": {
           "day": {
               "description": "Retrieves sales data for this day, formatted as YYYY-MM-DD.",
               "type": "str",
               "required": True
           }
       }
   }
]

Why Use Cohere for Tool Use?

Command R+ excels at leveraging external tools. This is evidenced by the model’s strong performance on the Microsoft ToolTalk benchmark, which evaluates a model’s ability to utilize tools and APIs. Command R+ strikes a great balance between performance and cost-efficiency. Command R+ achieves tools performance that is competitive with Claude 3 Opus and GPT-4 turbo while maintaining a three to five times lower total cost of ownership.

We evaluated conversational tool use performance using Microsoft’s ToolTalk (Hard) benchmark, and the cost is based on running the ToolTalk benchmark on the different models. Predicted tool calls were evaluated against the ground-truth, with overall conversation success metrics deemed on how likely the model recalls all tool calls and avoids bad actions (i.e., a tool call that has unwanted side effects).

At Cohere, we spent extra effort teaching our models to use enterprise tools. We began by training models to leverage vector databases and information retrieval systems to solve retrieval-augmented generation (RAG) use cases. This has naturally evolved into broader tool use — training the model to use any tool.

Finally, a unique advantage of Cohere's tools offering is the model’s transparency and verifiability. Command R and R+ models only generate claims that are verifiable through fine-grained citations. These models explain which tool results were used to generate every response, providing visibility into the model’s reasoning.

Get Started with Tool Use

Jump directly into the code with this example notebook, where we build a data analyst agent that has access to two tools: web search and a Python interpreter. This AI agent, once deployed, can become accessible to everyone in your organization, regardless of their technical background. Commands are simply natural language! Imagine extending data analysis capabilities to your entire team, without needing technical expertise.

For example, the AI agent built in the notebook will be able to answer the following request:

"Create a plot of the number of full time employees at the 3 tech companies with the highest market cap in the United States in 2024."

0:00

/0:36

To build your own AI agent, interacting with your own enterprise tools, check out our multi-step tool use documentation. The tools capabilities are available through our API and through Langchain.