The Receipt Machine
Every station of the workshop
Each station builds one piece of a retrieval pipeline: load, chunk, embed, retrieve, answer. You run it in your own AI and paste back the result to be scored. This list shows only stations that passed every check today.
- The X-ray: watch your AI find things
- Station 1, Load: get clean text out of a messy export Retrieval can only find what loading kept. A flattened table loses its meaning unless you rebuild it.
- Station 2, Chunk: cut it where the meaning lives A chunk that cuts a sentence in half can never be the receipt for it.
- Station 3, Embed: find passages that mean the same thing Embeddings put passages that mean the same thing close together, even when they share no words.
- Station 4, Retrieve: pull the right passage and catch the trap The passage that shares the most words with the question is not always the one that answers it.
- The miss lab: when the retriever finds nothing A word-matching retriever misses "home" when the handbook says "remotely". Rewrite the question until it hits.
- Station 5, Answer: every sentence with a receipt The last station: answer only from the sources, cite every sentence, and say so when the sources are silent.
- Capstone: your pipeline, on your own documents Write the spec for your own pipeline: what goes in, how it is cut, how much is retrieved, and five questions that prove it works.
- Station 5, Answer: every sentence with a receipt The last station: answer only from the sources, cite every sentence, and say so when the sources are silent.
- Station 3, Chunk: keep sentences whole A chunk that splits a sentence cannot be verified, so every chunk must end where a sentence ends.