Standardizing AI Frameworks in Odoo (pydanticAI, LangChain, etc.)

[ayushin]

"I see that the 'pydanticAI' framework has not been used. Why? It could be a good starting link to normalize certain things." – @ValentinVinagre

There are multiple ways to structure AI interactions in Odoo, from LangChain-based agents, llamaindex retrieval to pydanticAI for validation.

• Should we establish a standard framework to ensure modularity and interoperability?
• How do different tools compare in terms of usability, flexibility, and performance?
• Should we focus on one primary framework, or allow multiple integrations?

My point of view is that we should not tie Odoo LLM framework to a single AI framework and support each of them via extension modules.

[Kerrim-]

From our experience building MuK AI Studio (muk_ai) - which aligns with what you have planned - we’ve found that avoiding frameworks often leads to unnecessary complexity and redundant work.

In our development, we initially started with Gemini and later integrated ChatGPT. Along the way, we encountered challenges such as standardizing AI provider interfaces and managing execution loops - exactly the kinds of problems that existing frameworks, like LangChain in our case, have already solved effectively.

That said, muk_ai is not as modular as what you envision. Our focus was not on supporting multiple frameworks or making them highly modular, but rather on implementing the functionalities we needed for customers in a structured way.

E.g. Adding tools for execution (not complete):

from langchain_core.tools import StructuredTool
from langchain.agents import AgentExecutor, create_tool_calling_agent

tools = []
# we loop through all available tools here (we implemented the tools as an Odoo model)
tools.append(StructuredTool.from_function(func=tool_fn, name=record.type, description=record.description,))
# and pass them to an executor (without needing to know which model will execute them)
agent = create_tool_calling_agent(model, tools, prompt)
AgentExecutor(agent=agent, tools=tools, verbose=False, stream_runnable=False,)

[Kerrim-]

In the meantime I also checked out your PR regarding tool execution. I like the fact you have a consent workflow before execution. We solved that by giving the customer a customization screen for models where you can enable or disable access or use a domain to limit it.

In the PR you can see that implementing tool execution for all llms without a framework will be some work. and if something changes you will have to update a lot of files. Just my 2 cents :)