agentami 0.1.7

Create an agent that can handle a large number of tools with persistence support.

AgentAmi

AgentAmi is a flexible agentic framework built using LangGraph, designed to scale with large numbers of tools and intelligently select the most relevant ones for a given user query.

It supports:

• Dynamic tool selection via inbuilt RAG with an option to easily integrate your own tool_selector.
• State checkpointing with pluggable backends
• Pruner to limit memory and improve performance

---

Quick start

pip install agentami
from agentami import AgentAmi

Usage and features

Refer the main.py file for a complete sample usage.

from langchain.chat_models import ChatOpenAI
from langgraph.checkpoint.memory import InMemorySaver
from agentami.agents.ami import AgentAmi

# Replace ... (ellipsis) with the commented instructions

tools = [...]  # List of LangChain-compatible tools
agent = AgentAmi(
    model=ChatOpenAI(model="gpt-4o"),
    tools=tools,  # List of LangChain-compatible tools
    checkpointer=InMemorySaver(),  # Optional. No persistence if omitted.

    # Optional parameters:
    tool_selector=...,  # Custom function to select relevant tools. Defaults to internal tool_selector given.
    top_k=...,  # Number of top tools to use. Defaults to 3.
    context_size=...,  # Number of past user prompts to retain. Defaults to 7.
    disable_pruner=...,  # If True, disables pruning. May increase token usage. Defaults to False
    prompt_template=...  # Custom prompt template. Defaults to a generic bot template.
)
agent_ami = agent.graph # Your regular langgraph's graph.

How to integrate your own tool selector?

Just make a function that accepts (query: str, top_k: int) and parameters and returns List[str] #List of tool names.

from typing import List


# function template:
def my_own_tool_selector(query: str, top_k: int) -> List[str]:
    # Your logic to select tools based on the query
    return ["tool1", "tool2", "tool3"]  # Return top_k selected tool names

Things you should be aware about:

• Running for the first time will take time as it installs the dependencies (models used by internal tool_selector).
• Your first agent_ami.invoke() or agent_agent_ami.astream() may take time if you have hundreds of tools, because it initialises a vector store and embeds the tool descriptions at runtime for each AgentAmi() object
• Your eventual prompts would be fine.