7. ReAct: Combining reasoning and action in LLMs

Get Started. It's Free
or sign up with your email address
7. ReAct: Combining reasoning and action in LLMs by Mind Map: 7. ReAct: Combining reasoning and action in LLMs

1. Most applications require more complex workflows than a PAL model, with relations to multiple external data sources and apps

1.1. You can use ReAct framework to help LLMs plan out and execute workflows

1.1.1. ReAct: Synergizing Reasoning and Action in LLMs

1.1.1.1. A prompting strategy

1.1.1.1.1. combines

1.1.1.2. Uses one or few shot inference to show a LLM how to reason through a problem and decide a plan of actions to get closer to a solution

1.1.1.2.1. The shot includes

1.1.1.3. Instructions to define action space

2. Frameworks for developing applications powered by LLM are in active development

2.1. LangChain

2.1.1. Provides modular pieces that contain necessary components to work with LLMs

2.1.1.1. Prompt templates

2.1.1.2. Memory

2.1.1.2.1. To store interactions with a LLM

2.1.1.3. Pre-built tools

2.1.1.3.1. Enable developers to carry out a wide variety of tasks

2.1.1.4. Those three components together result in a chain

2.1.1.4.1. LangChain has set of predefined chains optimized for various use case and quick to use

2.1.1.4.2. But when application workflow varies on the information provided by the user, can't use pre determined chain

2.1.1.5. Agent

2.1.1.5.1. Used to interpret input from user and determine which tool(s) to use for task completion

2.1.1.5.2. LangChain includes agents for ReAct and PAL, as well as many others

2.1.1.5.3. Agents can be incorporated into chains to take action or plan actions and execute a series of them