Context Budget Assembler
An AI can only read so much at once. See what it reads, and what gets left behind, when you make its space bigger or smaller.
New here? Read the short guide for this demo, with steps to follow and what to look for.
The idea in one minute
Think of it as a suitcase.
Before an AI answers, we hand it things to read: the customer's question, tools it can use, the earlier chat and help articles. But it can only read so much at once. If we pack everything, it is slow, costly and easily confused. If we pack too little, it lacks the facts.
The context assembler is the packer. It keeps what matters for this question and leaves the rest behind. Try it below.
- 1Choose a question a customer might ask.
- 2Make the suitcase bigger or smaller with the slider.
- 3See what is packed and what is left out.
This is a demo with made-up data, and it runs only in your browser. The shop, the chat, the tools and the help articles are invented samples. No AI model is called, nothing you type is sent anywhere, and nothing is stored. Sizes are measured in tokens, which are small chunks of text, roughly three quarters of a word each. Here they are estimated at about 4 characters per token.
1 The customer asks
2 Size of the suitcase
3 How the room is used
The AI always gets its instructions and the question. A slice is also kept back so it has room to write its answer.
Tools it could use
Actions the AI may take. Each one costs space to describe.
The earlier chat
What was said before this question.
Help articles
Facts the answer can be based on.
See the exact text the AI would receive
Only the packed items, in a fixed order. This is what a developer would send to the model.
How the packer decides
Four simple rules.
- Always pack the instructions and the customer's question.
- Keep a slice free so the AI has room to write its answer.
- Score every tool, past message and article by how related it is to the question. Leave out the unrelated ones.
- Pack the best-scoring items first. When the suitcase is full, the rest are left out.
Honest limits
What this demo does not do.
- Relevance here is simple word matching. A real system would use embeddings and a re-ranker.
- Sizes are estimated from text length. A real system uses the model's own tokenizer.
- Old messages are dropped, not summarised. A real system would fold them into the summary.
Have an idea? Let's build it properly.
Tell us what you want to build or automate. We'll reply with practical next steps.
