Refactor to multi-agent system -> KGB agent

[anngvu]

We currently have 6 tools but can expect to have between 10-20. Soon it would be desirable to refactor to a multi-agent setup for performance, scalability, and improved modularity. See https://platform.openai.com/docs/guides/function-calling/keep-the-number-of-functions-low-for-higher-accuracy

We recommend that you use no more than 20 functions in a single tool call. Developers typically see a reduction in the model’s ability to select the correct tool once they have between 10-20 tools.

If your use case requires the model to be able to pick between a large number of functions, you may want to explore fine-tuning (learn more) or break out the tools and group them logically to create a multi-agent system.

The current tools can be grouped into "Synapse agent" and "knowledge graph RAG agent".

Update: This will be scoped to Knowledge Graph Bot agent and a separate issue planned for Synapse agent.

[anngvu]

Aside for the modularity benefits and practical performance reason mentioned above, in the long run it may be better than having one same model with many tools/hats because:

• We will likely want to plug in different trained models for each "agent". For example, the Knowlege Graph RAG agent may be a local model specifically trained on Datalog queries, while the general Extraction Agent is still an OpenAI model on the smarter side.
• Allows some calls to be cheaper and faster on "multi-threaded" mode. We can offload a task to another agent with just the instructions needed and not have to worry about intermediate attempts taking up precious context since only the final result is added back to the conversation (for example, AI models often have to iteratively refine SQL/SPARQL queries a number of times before succeeding). However, there is coordination overhead and loss of context can lead to poorer results in some cases.

However, there are a couple of things to think about:

• How and when do we want to give an agent context? For example, the extraction agent is called without the full message history, only the directive to extract.