OrDora/coachtwin-embedder

CoachTwin Embedder

The sentence embedding model powering workout retrieval in the CoachTwin app. It is BAAI/bge-small-en-v1.5, chosen by evaluating three encoders on the CoachTwin Workouts dataset (10,393 workouts).

Why this model

Leave-one-out retrieval. Strict relevance requires a match on both goal and body_focus; loose requires body_focus.

model	params	dim	strict P@3	loose P@3	MRR@10	corpus encode
bge-small-en-v1.5	33M	384	0.6687	0.8273	0.8020	10.4s
all-mpnet-base-v2	110M	768	0.5047	0.7453	0.6866	31.2s
all-MiniLM-L6-v2	22M	384	0.4733	0.6667	0.6686	7.0s

Random-retrieval baseline: strict P@3 0.0220, loose 0.1200, MRR@10 0.0685. A precision number without its baseline is not interpretable.

Selected: BAAI/bge-small-en-v1.5 - strict P@3 0.669, about 30x random.

The middle row is the interesting one: all-mpnet-base-v2 is 3.3x the parameters, 3x slower, and scores worse. The bigger encoder is not the better one here.

Usage

from sentence_transformers import SentenceTransformer
model = SentenceTransformer("OrDora/coachtwin-embedder")
emb = model.encode([text], normalize_embeddings=True)

Documents and queries use different templates, both recorded in embedding_info.json. This is a BGE model, so queries - not documents - take the prefix Represent this sentence for searching relevant passages: (needs_query_prefix: true).

Serialization note

Saved in the sentence-transformers 3.x module format. A repo saved by 5.x fails on 3.x with No module named 'sentence_transformers.base', and a client with a try/except fallback then silently swaps in a different encoder - no error, wrong neighbours, because several candidates share 384 dimensions.

Limitations

Base checkpoint, not fine-tuned. Evaluated only on synthetic English workout descriptions. Not fitness or medical advice.