Module Ensemble

The ModuleEnsemble Cog enables ensembling methods within generative AI workflows, allowing multiple independent sampler to generate diverse outputs that are subsequently aggregated to produce a final, refined response. This approach enhances the robustness and quality of generated content by leveraging the collective strengths of various solution path.

Implementation

All options within the ModuleEnsemble Cog follow a common process: sample then aggregate. To facilitate this process, the option transforms the target module into a workflow, wherein the entry point is multiple independent samplers each initialized with a clone of the original module, followed by an aggregator that processes proposals of these samplers and to generate a final response.

Note

The ensemble cog introduces significant structural change to the whole workflow, adding multiple modules that can be tuned independently. Take self-consistency as an example, choosing a different configuration for each sampling path can lead to a more diverse set of responses, increasing the chance of hitting the correct answer and reducing output bias.

Universal-Self-Consistency

The UniversalSelfConsistency option provides a simple implementation of the paper. It operates by spawning multiple workers and adding a llm-based aggregator. The aggregator analyzes all proposals and synthesizes a final answer that reflects the majority consensus. This method is particularly effective for tasks requiring complex reasoning and ensures that the final output is both comprehensive and coherent.

Example Usage

To utilize the ModuleEnsemble Cog with the UniversalSelfConsistency option, you can define the search space as follows:

from cognify.hub.cogs import NoChange, ModuleEnsemble, UniversalSelfConsistency

# Define the ensemble options
ensemble_options = [
   NoChange(),
   # We can also ask the option to change the temperature during sampling
   UniversalSelfConsistency(num_path=5, temperature=0.7, change_temperature=True)
]

# Initialize the ModuleEnsemble Cog
ensemble_cog = ModuleEnsemble(
   name="ensemble_example",
   options=ensemble_options,
)

Cognify also allows direct application of any ensemble options to your workflow:

import cognify
from cognify.graph.program import Workflow
from cognify.hub.cogs import UniversalSelfConsistency

agent = cognify.Model(...)
usc = UniversalSelfConsistency(num_path=3)
ensembled_new_agent: Workflow = usc.apply(agent)

# Easiest way to use the workflow with the same interface
agent.invoke = ensembled_new_agent.invoke
agent.forward(
   {
      "input_label_1": "value ...",
      "input_label_2": "value ..."
   }
)