Welcome — start here, and how this place works

[akash-coded]

Welcome to the Advanced RAG playground for cohorts C1 and C2.

First 30 minutes

1. Clone, make setup, make lab
2. Open notebooks/00_start_here.ipynb, press Run All
3. When it finishes, read the honesty inventory at the bottom — the table of what is real and
   what is a stand-in. Everything else in the curriculum depends on you having read it.

Where things go

You want to…	Go to
Ask why something behaves the way it does	Q&A — not Issues
Get an architecture challenged before building it	Design Reviews
Show what you built, including what did not work	Show & Tell
Submit an exercise	An exercise issue + a branch
Report a genuine defect with a reproduction	A bug issue
Practise an interview answer	Interview Prep

Three house rules

1 · Evidence beats adjectives. "Recall seems low" is not answerable. "Evidence Recall@8 is
0.61 on the temporal slice and 0.79 elsewhere" is.

2 · Every claimed improvement carries an interval. metrics.paired_bootstrap gives you one.
A delta inside the noise band is reported as inside the noise band — not rounded into a win.
This is enforced by the PR template and by CI.

3 · Negative results are full credit. "I tried HyDE and it did not clear the noise band,
here is the mechanism I think explains that" is one of the most useful things you can post.
The repository itself reports three findings that contradict the source material — see the
README.

A thing worth knowing on day one

There is a good first issue label with real, scoped work behind it, and a docs/EXERCISES.md
with 22 exercises graded by difficulty. You are not expected to invent your own starting point.

Ask anything. The only bad question here is the one asked in a DM, because it helps one person
instead of every future cohort.

[akash-coded]

Dana · reviewer · owns the eval harness

One thing worth adding for people arriving from a search or platform background rather than an ML one.

The instinct here is to read the toolkit first — store.py, retrieve.py — because that is where the code you recognise lives. Read notebook 01 first anyway. The whole repository is organised around the idea that a number without an interval is an anecdote, and if you skip the measurement material the retrieval code looks over-engineered — you cannot see what the harness is protecting against.

Fastest path if you have an hour: notebook 01 end to end, then §1.3 twice.

[akash-coded]

Priya · cohort peer

Practical note for anyone who hits it: make lab on a fresh machine wants the dev extras, and pip install -e . alone is not enough — you want make setup or pip install -e ".[dev]". Cost me ten minutes before I read the Makefile.

Also worth saying out loud since it took me a while to believe it: there really is no dataset download and no API key. If you are waiting for something to install, you have gone wrong somewhere.