Different LLM workflows to build a system
- Prompt chaining - Task is broken into a sequence of steps and each LLM call processes output of previous one.
- Routing - Router classifies an input and sends it through one of many predefined routes.
- Parallelisation - Bunch of LLMs work in parallel to do things and their output is aggregated. Could be good for voting mechanism for evaluation.
- Orchestrator-workers - Central LLM breaks down tasks and delegates them to worker LLMs.
- Evaluator optimiser - LLM generates a response gets feedback and then generates a response again.
- Agents - No fixed path. LLMs do their thing.
---
(Building Effective AI Agents, n.d.)