The missing sentence / a RAG experiment

Find the sentence
the answer missed

You want to bring a candle lantern on Saturday's garden walk. Search the guide and check whether the answer includes the Saturday restriction.

Your challenge: rescue the missing restriction with a 24-word evidence budget.

A document search can find the right guide while leaving out its exception. Change the search or passage budget and inspect the packed text. This helps you diagnose whether a missing condition was lost before an answer was written.

A short, two-attempt experiment. No account or API key.

This page really searches and chooses which passages fit. Its number-based search uses hand-set values, and its answer follows a written template. No language model is answering.

Moonlight Garden, with an unlit lantern waiting at its gate
Moonlight Garden / fictional visitor guide
What actually runs, and why it matters

Retrieval-augmented generation (RAG) adds retrieved evidence to a model's input. This lesson isolates the retrieval and context steps. It uses original fictional sources and an authored answer interpreter, not recorded or live model output. AI assisted the development; Jonathan Simone directs this independent project.

  • Keyword: count occurrences of question terms after lowercasing and removing the listed common words.
  • Meaning: cosine similarity over three hand-written coordinates. These are teaching vectors, not learned embeddings. Joining passages averages their vectors.
  • Combined: reciprocal-rank fusion, 1/(60 + keyword rank) + 1/(60 + meaning rank). Scores from different methods are not directly comparable. Ties sort by passage ID.
  • Packing: visit passages in rank order, or your chosen first passage. Include a whole passage if it fits; otherwise skip it and keep looking. Headers and metadata do not count toward the teaching word budget.
  • Support: explicit fixture labels connect a rule, its exception and a declared correction. Similarity does not establish truth. A larger context does not guarantee a better answer.

Technical sources: Lewis and colleagues on RAG, Microsoft on reciprocal-rank fusion, and the Stanford introduction to term frequency.

This optional check-in and your explanation preference are saved in your browser. Your exercise stays on this page unless you download it. View or clear course check-ins.