Tool Learning with Foundation Models

tags:: #source/article Large Language Models LLM Agents
Source

Source code and dataset: https://github.com/OpenBMB/BMTools

• Python code for running LLM's with tools (particularly OpenAI-Plugins)
• Uses Langchain

Example of Langchain Structured Chatprompt

Prefix

Respond to the human as helpfully and accurately as possible. You have access to the following tools:

Format Instructions

Use a json blob to specify a tool by providing an action key (tool name) and an action_input key (tool input).

Valid "action" values: "Final Answer" or

Provide only ONE action per $JSON_BLOB, as shown:

{
  "action": $TOOL_NAME,
  // ... other action parameters
}

Follow this format:

Question: input question to answer
Thought: consider previous and subsequent steps
Action:

$JSON_BLOB

Observation: action result
... (repeat Thought/Action/Observation N times)
Thought: I know what to respond
Action:

{
  "action": "Final Answer",
  "action_input": "Final response to human"
}

Suffix

Begin! Reminder to ALWAYS respond with a valid json blob of a single action. Use tools if necessary. Respond directly if appropriate.

Format is Action:

$JSON_BLOB

then Observation:.
Thought:

Prompt Structure

• System
  • Prefix
  • Formatted tools
  • Format instructions
  • Suffix
• Memory Prompts
• User

Gotchas

gpt turbo frequently ignores the directive to emit a single action