Search is just the library counter — reading is still your job. Answer without ever opening the book you were handed, and you're back to guessing.
Let's think about it this way for a second
Most RAG bugs don't come from bad search — they come from search results nobody actually read. Get back the top 3 chunks, and you (or the model) still need to read them and toss out whatever isn't relevant. Don't assume "the top result must be right." The library just hands you three books off a nearby shelf — you only find out which one actually answers your question once you read it.
Let's connect it to everyday life
Say the question is about phones, and the search returns: 1) phone policy, 2) computer usage, 3) cafeteria rules. Read through them, keep only #1, and discard #2 and #3 — then answer. In code, before you feed retrieved text into the prompt, check "does this actually match the question?" If it doesn't, respond with "not found."
Let's try it hands-on together
Retrieved:
1. phones-in-class.md ← keep
2. computer-lab.md ← drop
3. cafeteria.md ← drop
Only #1 goes into the answer prompt.You'll be able to sort search results into keep or discard.5-Minute Try-It
Write one question along with 4 sample chunks. Decide which to keep, which to discard, and explain why.
One Quick Warning
The noisier the chunks you feed it, the more the model tends to fill gaps with guesses. Feed it a little, but keep it clean.