Advice on multi-agent orchistration

[liefra]

I'm looking for advice on how to implement a multi-agent bot.

The idea is as follows:

1. A user's message comes in and gets classified using the excellent EmbeddingModelTextClassifier.
2. Depending on the classification result, the user's message is routed to one of several AiServices agents.
3. The AiServices agent processes the user's message using its specific set of retrievers and tools.
4. The result of the processing is added to the user's chat history, which contains the all history, spanning all agents.

The reason multiple agents are needed is that each of them can have many tools, and it would not be efficient enough to include all tools in every chat request. Agents can also be added and removed dynamically depending on additional user preferences.

Do you think it would be possible to implement something like this using LangChain4j (v0.22.0)?

Any advice would be much appreciated. Thanks!

[langchain4j]

@liefra you can try setting up multiple AI Services each having it's own set of tools and ChatMemory/ChatMemoryProvider. When the request comes in, you can classify it and route to the appropriate agent. But maybe this can be done simpler by having a set of tools selected dynamically for each request. Say if user asks math question, only math-related tools are provided to the LLM. (This is not possible yet.) What do you think?

[liefra]

Hi @langchain4j: If it were possible to dynamically select a set of tools for each request, that would be excellent. I believe this approach would be a nice balance between simplicity and the potential for scaling up an AiServices agent's complexity. For our use case, we could determine the set of tools based on the classification results.

Do you think adding this dynamic tool assignment per request would be feasible?