naas.ai Docs

LocalModel

3 min read

AirgapChatOpenAI

What it is

  • AirgapChatOpenAI is a small subclass of langchain_openai.ChatOpenAI that:
    • Builds a single prompt combining any SystemMessage content with the latest HumanMessage.
    • Adds a plain-text “tool list” to the prompt when tools are bound.
    • Parses model output for simple tool calls (TOOL_CALL: ...) and (optionally) runs the tools, then re-prompts the model with tool results.

Public API

Class: AirgapChatOpenAI(ChatOpenAI)

  • Purpose: Minimal wrapper around ChatOpenAI with basic, prompt-based tool support.

__init__(**kwargs)

  • Pass-through to ChatOpenAI.__init__.
  • Initializes an internal tool registry (self._tools = []).

bind_tools(tools, **kwargs)

  • Stores tools in self._tools and returns self.
  • Tools are expected (by convention) to expose:
    • name and description attributes (for prompt listing)
    • Either an invoke(args: dict) method or be callable (tool(**args))

bind(**kwargs)

  • Calls ChatOpenAI.bind(**clean_kwargs) but removes any kwargs whose key contains "tool" (case-insensitive).
  • If no supported kwargs remain, returns self.

_generate(messages, stop=None, run_manager=None, **kwargs) -> ChatResult

  • Overrides ChatOpenAI._generate.
  • Behavior:
    • Extracts and concatenates all system-message content strings into a system_prompt.
    • Uses the last HumanMessage content as the user message.
    • If tools are bound, appends:
      • A list of tools (- {name}: {description})
      • Instruction: To use a tool, respond with: TOOL_CALL: tool_name {json_args}
    • Replaces messages with a single HumanMessage containing the constructed prompt.
    • Filters generation kwargs to only: temperature, max_tokens, stop.
    • After first generation, if tools are bound:
      • Scans the model output for TOOL_CALL:\s*(\w+)\s*({.*?}) (DOTALL)
      • For each match:
        • Parses JSON args
        • Finds tool by tool.name
        • Executes via tool.invoke(args) if present, else tool(**args)
        • Collects results/errors
      • If any tool results exist, re-calls the model with:
        • Original content
        • Tool results: block
        • Provide a final response:

Configuration/Dependencies

  • Requires:
    • langchain_openai.ChatOpenAI
    • langchain_core message/output/callback types
  • Uses standard library:
    • json, re
  • Runtime configuration is inherited from ChatOpenAI (e.g., model, API settings, etc.).

Usage

from langchain_core.messages import HumanMessage
from naas_abi_core.services.agent.beta.LocalModel import AirgapChatOpenAI
 
class EchoTool:
    name = "echo"
    description = "Echo back the provided text."
    def invoke(self, args):
        return args["text"]
 
llm = AirgapChatOpenAI(model="gpt-4o-mini", temperature=0).bind_tools([EchoTool()])
 
# Ask the model; if it emits "TOOL_CALL: echo {...}", the tool will run and a final answer is requested.
result = llm._generate([HumanMessage(content="Use the echo tool with text='hello' and then answer.")])
print(result.generations[0].message.content)

Caveats

  • Tool calling is prompt-based and regex-parsed, not using OpenAI/LC structured tool-call APIs.
  • Only tool names matching \w+ are recognized, and JSON args must match {.*?} in the output.
  • Only the last HumanMessage is used; multiple user turns are not preserved as separate messages.
  • _generate is an internal LangChain method; depending on your LangChain usage, you may prefer calling higher-level interfaces that eventually delegate to _generate.

Graph View