Six months ago we set out to build the best semantic code search platform we could: a RAG pipeline that gives LLM agents precise, citable evidence from real repositories instead of whatever grep happens to surface. We got it working, we got it into production, and we collected a lot of scar tissue along the way. This talk is us sharing the parts we wish someone had told us on day one.

In the session we'll walk through the steps of building a RAG pipeline, what we learned, and which mistakes to avoid. How a retrieved chunk without coordinates is almost useless, and metadata quietly ends up being half of your retrieval quality. How binarizing embeddings down to one bit per dimension costs only about 2% retrieval quality while slashing storage and latency, which still feels wrong every time we say it. How precision barely matters in code retrieval while recall is everything. And how agents, go figure, have personalities: Claude is deeply suspicious of external tools and will quietly go back to grep instead of the search we built for it, so we spent real time reading trajectories and measuring invocation rates like behavioral psychologists.

We'll also touch on ideas we borrowed from RAPTOR and LLM2Vec, why we stopped trusting SWE-bench Verified, and how we keep the index breathing with git using content hashes and merkle trees instead of reindexing the world on every commit.

If you're building anything like this, you'll leave with our concrete recipe, real model and tool choices included, from laptop prototype to production setup.

This session targets a gap many AI builders are facing: agent quality, cost, and latency are dominated by what lands in the context window, not by which model they picked. Retrieval over source code is the lever, and it's poorly covered.

What makes this session valuable is that it's a practitioner's report, not a survey. We spent six months building a semantic code search platform and shipping it to production, and every claim comes with a decision and a number attached: which embedding and reranker models we run, why binary quantization was nearly free, why we abandoned SWE-bench Verified, how we measured whether the agent actually used the tool. Attendees leave with a recipe they can start applying the same week, at either prototype or production scale.

It fits an AI builders track well because it sits exactly where the audience works: between models they don't control and products they have to ship. It's technical enough for infrastructure engineers, but the lessons about evals, agent behavior, and context quality apply to anyone building on LLMs.