RAG, without the magic
A non-technical path from deciding whether retrieval belongs in the system to measuring which half failed.
Eight progressively connected essays on retrieval-augmented generation: the decision, the pipeline, document boundaries, semantic representations, relevance, query intent, architecture, and evaluation.
Should this be RAG?
Decide whether the missing piece is knowledge, behavior, or both.
What happens before the answer?
One loop prepares the library. Another finds the evidence.
The boundary becomes part of the answer
Choose what stays together before retrieval decides what matters.
How does a search find the same idea in different words?
Embeddings estimate relatedness. Indexes retrieve candidates.
Similar is not relevant
Related passages are candidates. The answer still has to earn its place.
The retriever got the wrong question
Clarify, rewrite, split, or chain before retrieval.
The pipeline is part of the answer
Start fixed. Branch only when failure is measured.
The answer failed. Which half broke?
Evaluate both stages. Keep the trace.