Split Test AI Prompts Using Supabase & Langchain Agent

grawenbh — Sat, 28 Mar 2026 15:12:04 +0000

This workflow allows you to A/B test different prompts for an AI chatbot powered by Langchain and OpenAI. It uses Supabase to persist session state and randomly assigns users to either a baseline or alternative prompt, ensuring consistent prompt usage across the conversation.

🧠 Use Case Prompt optimization is crucial for maximizing the performance of AI assistants. This workflow helps you run controlled experiments on different prompt versions, giving you a reliable way to compare performance over time.

⚙️ How It Works
When a message is received, the system checks whether the session already exists in the Supabase table. If not, it randomly assigns the session to either the baseline or alternative prompt. The selected prompt is passed into a Langchain Agent using the OpenAI Chat Model. Postgres is used as chat memory for multi-turn conversation support.

🧪 Features

Randomized A/B split test per session
Supabase database for session persistence
Langchain Agent + OpenAI GPT-4o integration
PostgreSQL memory for maintaining chat context
Fully documented with sticky notes

🛠️ Setup Instructions Create a Supabase table named split_test_sessions with the following columns:

session_id (text)
show_alternative (boolean)

Add credentials for:

Supabase
OpenAI
PostgreSQL (for chat memory)

Modify the “Define Path Values” node to set your baseline and alternative prompts.

Activate the workflow.

Send messages to test both prompt paths in action.

🔄 Next Steps

Add tracking for conversions or feedback scores to compare outcomes by banana ai.
Modify the prompt content or model settings (e.g. temperature, model version).
Expand to multi-variant tests beyond A/B.

what is seedance 2 's advantage?

grawenbh — Wed, 18 Feb 2026 02:26:23 +0000

In the film and TV industry, storyboarding is extremely important, so there’s a long‑standing learning method called “shot breakdown.” It’s all about studying camera blocking, composition, and how emotions are guided. Learning the shot design of great masters and great works is very important.

In the past, if we wanted to imitate a certain style of work, it was very difficult. But now, because Seedance 2.0 supports video references, this has become much easier. For example, I really like this storyboard sequence from Weathering With You. Now all we need to do is feed that clip into the system as a reference and have it generate a brand‑new story for us – it’s incredibly convenient.

And it’s not just for narrative films and storyboards. You can use it in commercials as well. For instance, I can directly take a car commercial’s shot sequence and camera moves, combine it with a single image from DJI, and have it reproduce that style. Now anyone can use just one image to create a “million‑dollar” blockbuster.

One more note here: if you’re using the Doubao internal beta version, video reference upload is not supported yet. Presumably the Doubao user base is just too large, so you’ll have to wait a bit longer.

DEV Community: grawenbh

Split Test AI Prompts Using Supabase & Langchain Agent

what is seedance 2 's advantage?