This weekend, I enjoyed the opportunity to play with LlamaIndex Workflows at the LlamaIndex Ragathon thanks to Mark Evans, AI Makerspace and 500 Global.
Toolhouse and Apify provided easy, out-of-the-box connectors for the function calling of websites in our project to retrieve data. With such quick, easy and reliable setup, I got to step away from web-browsing automation to explore using LLM's for other tasks such as categorization, recommendation and reasoning.
My takeaway working with LlamaIndex Workflows is that it is brilliant to have a toolset to orchestrate function-calling and set up agentic architecture as event-driven systems.
It was extremely easy to comprehend and fast to implement a LlamaIndex Workflow, which is exactly the point of setting up an agentic, multi-LLM-step system - for prototyping purposes.
What this meant for me was that setting up an app that involved a recommendation system was able to be done in a matter of hours, as opposed to the complexity of traditional ML model tooling.
It was so easy to set up, my teammate Tanmesh Mishra and I each enjoyed our Saturdays and actually started working around 8pm Saturday night until the 3pm Sunday deadline, and we were still able to submit a working MVP with extra time to prepare slides.
I thought that having Event Classes to define what each step would do through its naming convention, and then a Workflow Class to orchestrate all the Event Classes as steps made the entire project extremely minimalistic.
Then, making use of Python's type hinting system as a pattern to define how one event would be mapped to another made code extremely organized.
Most of all, the ability of keeping the abstractions within a single file, with such clear structure allowed LLM code generation to be extremely successful.
As with all agent use-cases, having many LLM calls, especially if they are well-defined with examples and templated output makes agentic use-cases costly in money and time.
After having done this project, I would agree with the consensus of keeping the number of function-calling steps as minimal as possible when making multiple LLM requests, or at least having Human in the Loop intervention serving as checkpoints every few steps.
As more function-calls involving separate LLM requests stacks up, it makes debugging extremely time-consuming.
Here are all the steps in our workflow:
1. Tweet Extractor
2. Tweet Analyzer
3. Interest Mapper
4. Gift Idea Generator
5. Mediation
6. Gift Debater
7. Gift Reasoner
8. Amazon Keyword Generator
9. Amazon Product Link Generator
The above takes about 5 - 20 minutes to run start to finish, depending on the number of item categories and items chosen to debate over.
Here was the initial submission of GiftGenie:
https://lnkd.in/ghBZW2zB
Here is a second demo made after adding additional agent logic, Pro and Con agent dialogue over more gift choices, and a progress bar:
https://lnkd.in/gS_6n72A
Github Repo:
https://lnkd.in/gCNPHCKk
