Ground the Model: RAG Engineering Lab

09:00 – 10:15

Decide whether RAG is even the answer

when to use RAG and when not to, the failure modes, the alternatives

Takeaway Three real problems triaged: one to RAG, one to fine-tuning, one to a database query.

10:30 – 11:45

Ingest documents the way they actually arrive

parsing and cleaning, layout-aware extraction, messy PDFs

Takeaway A realistic document collection parsed clean, including the table that breaks naive extraction.

12:00 – 13:00

Chunk with intent

chunking strategies, semantic splitting, overlap and metadata

Takeaway The same corpus chunked three ways with retrieval quality compared, strategy chosen on evidence.

14:00 – 15:15

Embed and search by meaning

embeddings, vector search, similarity limits

Takeaway Semantic search running over your corpus, and one query where it loses to keyword search, explained.

15:30 – 17:00

Measure whether the right evidence comes back

retrieval evaluation, golden questions, recall vs precision

Takeaway An evaluation set that scores retrieval before any answer is generated: the day-two yardstick.

09:00 – 10:15

Generate answers that cite their evidence

grounded generation, citations, source attribution

Takeaway Answers with citations a reader can follow, and a refusal when the evidence is not there.

10:30 – 11:45

Add hybrid retrieval and reranking

hybrid search, filtering, rerankers

Takeaway A reranker added and measured, kept only because the evaluation set says it earns its cost.

12:00 – 13:00

Handle the failure cases

no evidence found, conflicting sources, stale documents

Takeaway The assistant behaving honestly on the three failure cases users actually hit.

14:00 – 15:15

Evaluate the whole pipeline

answer-quality evaluation, groundedness, regression checks

Takeaway End-to-end quality measured and a regression caught on purpose: quality as a number, not a feeling.

15:30 – 17:00

Improve it under time pressure

diagnosing weak answers, targeted fixes, the write-up

Takeaway The weakest answers diagnosed to their cause and fixed, with before-and-after scores you keep.