Building a retrieval pipeline that doesn't lie to you
Chunking, reranking, and evaluation for a RAG system that admits when it doesn't know.
Tools matter less than the feedback loop they enable. A fast, boring loop beat every sophisticated setup I tried, and I've kept it ever since, even as the rest of the stack changed twice over.
A short history
What surprised me most was how small the final change was. Most of the work was understanding the problem precisely enough that the fix became obvious, and that understanding took far longer than the fix itself.
By the end of this post you should be able to reproduce the whole thing yourself in under an hour, including the parts I got wrong the first time — those are marked so you can skip them if you're in a hurry.
The design
There is a version of this article that opens with theory. I decided against it, because the theory is easy to find and the motivation is not — so let's start with the thing that actually broke.
I want to be honest about the trade-offs here. The approach I settled on solves the problem I had, not every problem of this shape, and there is a clear point where you should reach for something more heavy-duty.
A note for context that helps the reader.
CREATE INDEX CONCURRENTLY idx_orders_created_at
ON orders (created_at DESC)
WHERE status = 'open';Implementation notes
The first version worked, which is exactly why it was hard to see how wrong it was. Everything looked fine in the demo, and everything looked fine until traffic arrived on a Tuesday afternoon and the graph started doing something I'd never seen before.
I kept a log of every decision during this project, and reading it back a month later was humbling: the choices that aged well were the boring ones, and the clever ones all came with a footnote I hadn't written yet.
The best engineering advice I've received is unglamorous: make it measurable, make it reversible, and only then make it fast.
Failure modes
What surprised me most was how small the final change was. Most of the work was understanding the problem precisely enough that the fix became obvious, and that understanding took far longer than the fix itself.
By the end of this post you should be able to reproduce the whole thing yourself in under an hour, including the parts I got wrong the first time — those are marked so you can skip them if you're in a hurry.
| Approach | Best for | Watch out for |
|---|---|---|
| Baseline | Quick comparisons | No tail visibility |
| Instrumented | Finding the real hotspot | Overhead of the probes |
| Production | Truthful numbers | Harder to reproduce |
When not to use this
What surprised me most was how small the final change was. Most of the work was understanding the problem precisely enough that the fix became obvious, and that understanding took far longer than the fix itself.
The numbers below are from my own machine and my own workload. Treat them as a method, not a benchmark: the point is how the measurement was taken, not the exact figures it produced.
Something the reader should not miss.
Further reading
Tools matter less than the feedback loop they enable. A fast, boring loop beat every sophisticated setup I tried, and I've kept it ever since, even as the rest of the stack changed twice over.
I kept a log of every decision during this project, and reading it back a month later was humbling: the choices that aged well were the boring ones, and the clever ones all came with a footnote I hadn't written yet.
# rebuild only what changed
docker build \
--cache-from "$IMAGE:latest" \
--tag "$IMAGE:${GIT_SHA}" \
.A short history
Tools matter less than the feedback loop they enable. A fast, boring loop beat every sophisticated setup I tried, and I've kept it ever since, even as the rest of the stack changed twice over.
The first version worked, which is exactly why it was hard to see how wrong it was. Everything looked fine in the demo, and everything looked fine until traffic arrived on a Tuesday afternoon and the graph started doing something I'd never seen before.
- Latency at p50, because most users live there
- Latency at p99, because someone always lives there
- Error rate, because a fast failure is still a failure
- Saturation, because averages hide the cliff
The design
The first version worked, which is exactly why it was hard to see how wrong it was. Everything looked fine in the demo, and everything looked fine until traffic arrived on a Tuesday afternoon and the graph started doing something I'd never seen before.
The first version worked, which is exactly why it was hard to see how wrong it was. Everything looked fine in the demo, and everything looked fine until traffic arrived on a Tuesday afternoon and the graph started doing something I'd never seen before.
By the end of this post you should be able to reproduce the whole thing yourself in under an hour, including the parts I got wrong the first time — those are marked so you can skip them if you're in a hurry.
Wrap-up
If you remember one thing, make it this: chunking, reranking, and evaluation for a rag system that admits when it doesn't know compounds. The cost of getting it right early is a afternoon; the cost of getting it late is a migration.