Support Mixture of Agents

[josephykwang]

What problem or use case are you trying to solve? Mixture-of-Agents Enhances Large Language Model Capabilities show it can outperform gpt-4o.

Describe the UX of the solution you'd like None

Do you have thoughts on the technical implementation? the llama_index implementation can be found https://github.com/run-llama/llama_index/blob/main/llama-index-packs/llama-index-packs-mixture-of-agents/README.md?__s=awieapaifz8bnwv7cphv&utm_source=drip&utm_medium=email&utm_campaign=LlamaIndex+news+2024-06-18

note that depending on the SWE-agent task, we may need to tune prompts

Describe alternatives you've considered

Additional context

[mamoodi]

@josephykwang would this issue be the same as what you're talking about? https://github.com/OpenDevin/OpenDevin/issues/2075

[rezzie-rich]

Something similar to #2075 was mentioned earlier in #486

[rezzie-rich]

This specific issue seems more like a replacement for MoE for llm architecture. However, this can be replicated using an agent framework.

The workflow could be like:

User prompt -> extracter-agent(extracting the objective) -> generator-agent(generating the first draft) -> reviewer-agent (reviews the draft and highlights the key areas) -> rag-agent (retrieves relevant info) -> verifier-agent (verifies reviewer-agents output with rag-agents findings) -> generator-agent (generates the final output)

Using multiple task specific llms, as mentioned in #2075 and #486, can get a cheaper and better output than just using gpt-4o for all of them.

[neubig]

@mamoodi , I think "mixture of agents" is a very specific version of #2075 and #486, so we can probably leave this one open as it is more concrete than the other ones. This would be cool!

[rezzie-rich]

This specific issue seems more like a replacement for MoE for llm architecture. However, this can be replicated using an agent framework.

The workflow could be like:

User prompt -> extracter-agent(extracting the objective) -> generator-agent(generating the first draft) -> reviewer-agent (reviews the draft and highlights the key areas) -> rag-agent (retrieves relevant info) -> verifier-agent (verifies reviewer-agents output with rag-agents findings) -> generator-agent (generates the final output)

Using multiple task specific llms, as mentioned in #2075 and #486, can get a cheaper and better output than just using gpt-4o for all of them.

Upon revisiting, it seems I misunderstood the content the first time. This is actually a variant of #3151. wouldn't you agree? @josephykwang

[github-actions[bot]]

This issue is stale because it has been open for 30 days with no activity. Remove stale label or comment or this will be closed in 7 days.

[github-actions[bot]]

This issue was closed because it has been stalled for over 30 days with no activity.