Why the baseline fails

A naive retriever breaks on a few recurring question shapes. Under-specified questions give the embedder too little signal. Multi-hop questions ask about separate topics at once and retrieve chunks about either, never both. Paraphrased questions use different vocabulary than the documents, so lexical similarity fails even when semantic similarity should work.

Flashcards: Flashcards

Loading practice…

Where the baseline breaks

Each failure mode has a specific fix later in the workshop.

Context windows are finite and attention degrades as context grows, so dumping twenty mediocre chunks produces worse answers than five great ones. Increasing top_k also does not help when the relevant chunk is not in the top fifty either, which is exactly what happens on paraphrased and multi-hop questions. You need better queries, not more results per query.

Quiz: Quiz

Loading practice…

Checkpoint: Baseline pipeline checkpoint

Loading practice…