almador2002/tripalchemy-embeddings

🧠 TripAlchemy Embeddings — MiniLM-L6

The winning embedding model + precomputed vectors for the TripAlchemy recommender

Precomputed 384-dimensional embeddings for all 10,396 TripAlchemy experiences, produced by the model that won a rigorous 3-model comparison — plus the full report so the choice is reproducible.

---

🏆 The 3-model comparison

Three Hugging Face sentence-embedding models were evaluated with a self-consistency retrieval test: for a stratified sample, a query built only from each experience's category + vibe tags (never its title/description) is embedded, and we measure whether the model ranks the original experience in its own top-5 (Recall@5) and its reciprocal rank (MRR). Final pick = weighted score (50% Recall@5 + 30% MRR + 20% encode-speed).

model	hf_id	Recall@5	MRR	encode time	dim
MiniLM-L6 ⭐	sentence-transformers/all-MiniLM-L6-v2	0.0267	0.0242	34s	384
BGE-small	BAAI/bge-small-en-v1.5	0.05	0.0446	60s	384
E5-small	intfloat/e5-small-v2	0.06	0.0423	62s	384

Winner: MiniLM-L6 (sentence-transformers/all-MiniLM-L6-v2). At 10,396 documents the three models' retrieval scores were statistically close, so the tie-break went to encoding speed — MiniLM-L6 is ~2× faster, ideal for a live Space that loads these vectors at startup.

ℹ️ The absolute Recall@5 numbers are low by design — the query is intentionally stripped to category + vibe only, so retrieving the one exact experience among 10,396 is hard. The relative ranking is what selects the model; real query→relevant retrieval in the app is excellent.

📦 Files

file	what it is
winning_model.npz	ids + 384-D document vectors for all 10,396 experiences
comparison_report.json	full metrics + the query/doc prefixes the app reuses
embeddings_compare_hf.py	the comparison code (reproducible)

🔧 How the app uses it

The TripAlchemy Space downloads winning_model.npz at runtime, loads it into a FAISS IndexFlatIP, and blends embedding similarity with the category sliders:

final_score = α · cosine(query, doc) + (1 − α) · slider_match

🔁 Reproduce

pip install sentence-transformers
python embeddings_compare_hf.py --sample-size 300

---

Part of TripAlchemy · 📊 dataset · 🚀 live app