Embeddings find meaning and miss exact words; keyword search finds exact words and misses meaning. Real retrieval usually runs both and merges the results, because the questions that break one are the ones the other answers.
About 1 min of reading
Before this new lesson, one small question about the lesson you just finished. Recall it, do not look it up.
See the exact token miss.
The question
Before you run it, say out loud what it will print. Then open the answer.
The keyword hit
Before you run it, say out loud what it will print. Then open the answer.
If your questions contain codes, IDs, or names, keyword search is not optional. An embedding alone will quietly drop the exact match, and that drop is the failure mode to check for.
Run a list of real questions and record which found nothing. That miss rate is the number that decides whether to add the other search, and guessing it is how a system ships broken.
A part number, an error code, or a name is an exact token an embedding may rank low. A question phrased differently from the document is the one keyword search misses entirely.
Run both, then combine the lists so a chunk found by either is a candidate. That union is what keeps a part number searchable without losing the paraphrased questions.
A chunk found by both searches is stronger than one found by either. Combining the two scores into one ranking is the standard way to merge without hand tuned weights.
Each one is defined here on this site. Open the entry for the full definition and the other words it points at.