The pieces are known; the system is the order of them plus the places a step can quietly return nothing. Build it as small functions with one job each, so a bad answer points at the function that caused it instead of at the whole.
About 1 min of reading
Run the stages separately.
Build stage
Before you run it, say out loud what it will print. Then open the answer.
Run stage
Before you run it, say out loud what it will print. Then open the answer.
Before you write anything, put these lines back in the right order. Click a line to move it up one step. This support step is easier than writing code from scratch, and it is here to make the last example stick.
Read the first example one line at a time. A line that prints shows what appears on screen; the rest set things up quietly.
Each stage can return nothing: no chunks, no filter survivors, no candidates above the threshold. Each needs its own message, because a blank answer tells the reader nothing about which stage failed.
A script that runs one stage on a saved input is what makes a bug findable. When the whole pipeline is one call, a wrong answer has seven possible causes and no way to narrow them.
Chunk, embed, index, search, filter, rerank, prompt. Each takes the previous output and returns a list. A stage that returns an empty list is a bug you can see, which a single long function cannot give you.
The function returns the answer plus the passages it used and their scores. That is the difference between a system someone can debug and a system someone can only complain about.
Chunking and embedding the documents happens once, at build. Only the question is embedded at run. Mixing them makes every question slow and makes the stored vectors stale.
Each one is defined here on this site. Open the entry for the full definition and the other words it points at.