LLM Workflow Patterns

Let's see five key patterns in which LLM Workflows are implemented.

Prompt Chaining

This workflow is used when we can break down a complex task into a sequence of steps, and each step can be processed as individual LLM call using the output of the previous step.
Example:
  1. Multistep reasoning tasks such as summarize text, critique and then rewrite
  2. Code generation pipeline such as generate plan, write code, test code, explain output


Router

This workflow is used in cases where an agent classifies the intent/task type/domain of the input query and then delegates the request to another agent, tool or workflow. 
Example:
  1. Multimodal agents such as those with separate text processing and image processing LLMs
  2. Multidomain assistants such as processing questions related to legal vs finance vs medical domains

Parallelization

The workflow pattern is useful when a task can be broken down into subtasks which can be handled concurrently by multiple LLM or agents. The outputs of these multiple calls are then aggregated by an aggregator to return a synthesized result to the user.
Example:
  1. Processing multiple documents/images in parallel
  2. Generating and sending emails to multiple clients in parallel



Orchestration

In this workflow, an orchestrator agent plans subtasks and delegates them to multiple specialized subagents or models each with specialized knowledge or domain expertise.
Examples:
  1. A Blog writing agent with researcher, writer and quality-checker subagents
  2. Coding agent with an architect, developer, code reviewer, tester agents 

Evaluators and Reflect-Refine Loop

This workflow provides a loop where one agent/LLM generates the result, and another agent/LLM evaluates the result based on a set criteria. The loop continues until the criteria is fulfilled or loop runs up-to a threshold. This workflow promotes selft-reflection, optimization and iterative improvement.
Example:
  1. Agents that uses a planner and reviewer agents in loop
  2. Blog writer agents with style enhancer agent providing feedbacks


Conclusion

All these patterns are not exclusive to each other and not to be used in isolation. We can create sophisticated complex workflows by using multiple patterns together. For example, by combining router and orchestration workflow, we can have an orchestrator which caters to multi-domain customer support agent.

Comments

Post a Comment

Popular posts from this blog

Summary of Paper: RSync Algorithm

 Rsync Algorithm:  TR-CS-96-05.dvi (cmu.edu) RSync is a utility which is used to transfer files or folders from computer to another in unix based systems. The key advantage of Rsync is once the full data is transferred to destination, only the changed bytes are transferred next time onwards, saving time and network bandwidth. The above mentioned paper describes the algorithm used to find the changed bytes and how to file data is transferred from source and recreated at destination. Here is the summary. Instead of transferring complete file, divide the file into chunks of bytes and calculate its hashes. First time all the bytes are transferred to the destination and then for each chunk of bytes a weak rolling checksum (inspired by Adler-32 Checksum) and strong checksum of each block of bytes is calculated and shared with the source. At source, the combination of rolling hash and strong hash are used to find the indexes of chunks of bytes that are same as some block of bytes des...
Read more

Monitoring and Restarting Apache webserver Automatically using AWS CloudWatch

 When your webserver goes down, you need find out why the webserver went down and fix the issue. But even before you troubleshoot and fix the issue, you need to make sure your webserver is available to your users so they can continue using your website. There are lot of sophisticated solutions are there, such as failing over to a redundant web server, or creating new instance using AWS AutoScaling. However this solutions comes with their own cost and complexity and if you are small shop like us you may not be able to afford it. And sometimes the simplest solution is to restart the webserver, if you don't have a failover server or AutoScaling configured for your server. If your infrastructure is on AWS, you can use couple of AWS services to restart the webserver automatically without any manual intervention. You need to use AWS CloudWatch Logs, CloudWatch Metrics, CloudWatch Alarm and AWS SystemManager RunCommand. Note that if you are using some Metrics published directly by AWS ser...
Read more

Continuous Integration and Delivery pipeline using Bitbucket, AWS CodeBuild, AWS CodePipline and AWS CodeDeploy - Part 1

Image
In this post I am going to show you how to develop a continuous integration and continuous delivery pipeline using AWS CodeBuild, AWS CodePipline and AWS CodeDeploy. I will use bitbucket as our source repository. But any other repository such as Github or Gitlab also can be used. Ofcourse, certain steps may defer as AWS Code Pipeline can pull the code directly from Github, but not from bitbucket. So here are the steps: 1. Configure AWS CodeBuild to build the code by directly pulling from BitBucket and upload the build artifacts in S3. 2. Configure AWS CodeDeploy to pull the build package from S3(configured in above step) and deploy the application. 3. Configure AWS CodePipeline to get the build artifacts from S3 and deploy them using the AWS CodeDeploy application configured in step 2. This will be three part series, and in this part-1 we will see how to configure AWS CodeBuild. Part 2: Configuring AWS Code Deploy Part 3: Configuring AWS CodePipeLine Step 1: Confi...
Read more