Bio-Bite Retriever

This repository is the deployment snapshot of the embedding model selected for the Bio-Bite university project. It is based on intfloat/e5-small-v2 and is not fine-tuned. Publishing the snapshot makes the exact retriever used by the app explicit and gives the project its own Model repository while retaining clear attribution to the base model.

Intended use

The model embeds English descriptions of a user's training, sleep and recovery state for similarity search over the synthetic Bio-Bite recipe dataset. It is an educational prototype, not a medical or nutritional model.

Evaluation

The final v2 evaluation uses 72 balanced queries across six recovery categories. recovery_category and Nutritional_Need are excluded from the embedded corpus text to avoid label leakage. Reported metrics include precision@3 with a bootstrap 95% confidence interval, MRR, nDCG@10, per-category performance, latency and a TF-IDF baseline.

The final run selected E5 as the best neural retriever:

TF-IDF remained the strongest overall baseline on this synthetic, vocabulary- regular corpus. E5 was selected as the strongest embedding model for the semantic retrieval component; the project does not claim it beat TF-IDF.

This repository includes the following generated evaluation files:

• embedding_model_comparison_v2.csv
• embedding_eval_by_category_v2.csv
• embedding_eval_per_query_v2.csv
• recsys_config_v2.json

Limitations

• The recipe corpus is synthetic.
• Recovery labels are educational categories, not clinical diagnoses.
• HRV differs substantially between individuals; the app treats it as one signal alongside sleep, strain and semantic context.
• This snapshot does not add learned parameters beyond the base E5 model.

Base-model attribution

Base model: intfloat/e5-small-v2. See the base repository for its model details and license.