Sell Your Logic

AI Playground

See retrieval-augmented generation (RAG) step by step. Change one setting at a time, run again, and compare the two runs side by side.

Knowledge base: our 5 case studies Search by meaning: ready 10 of 10 runs left today
1. The question

Sign in or join free to write your own question and prompt.

2. Retrieval

How the notes are cut (chunking)

One idea or step per chunk, a few sentences. The usual sweet spot.

3. The model

Nothing you type is stored. Hosted models are run by Google Gemini or Groq. On Google's free plan, Google may use this text to improve its products, and people at Google may read some of it, so please do not share personal details.

Four experiments to try

Each one sets up the form; press Run, change the one setting it names, and run again.

What happens when you press Run. Your question becomes a vector of 768 numbers. Postgres finds the closest chunks of our case studies by cosine similarity. Chunks below the minimum similarity are dropped, and the best top-k become the notes in the prompt. Then the model answers, and you see every step, the time each took and a rough cost. Read the RAG case study.