tripmatch-ai-embedding-model

TripMatch AI Embedding Model

The production text encoder for the TripMatch AI Final Project. This repository is an attributed project copy of sentence-transformers/paraphrase-MiniLM-L3-v2, selected for semantic trip retrieval after comparing three Hugging Face embedding models on the same 10,000-row catalog.

Use

from sentence_transformers import SentenceTransformer

model = SentenceTransformer("avihayamor/tripmatch-ai-embedding-model")
vectors = model.encode(
    ["a seven-day scenic road trip"],
    normalize_embeddings=True,
)

The public Space loads this exact repository directly and ranks normalized catalog vectors with cosine similarity plus transparent structured compatibility.

Selection criteria

The final notebook and dataset repository publish the complete comparison of:

• sentence-transformers/all-MiniLM-L6-v2
• sentence-transformers/paraphrase-MiniLM-L3-v2 (this winner)
• sentence-transformers/all-MiniLM-L12-v2

The decision considers top-three constraint compatibility, query latency, full-catalog encoding time, dimension, and artifact size. MiniLM-L3 slightly improves the measured compatibility over L6 and is substantially faster than L12, making it the Pareto choice for a public CPU-Basic Gradio Space.

Data and evaluation

• Dataset: https://huggingface.co/datasets/avihayamor/tripmatch-ai-dataset
• Application: https://huggingface.co/spaces/avihayamor/tripmatch-ai
• Catalog: 10,000 English synthetic trip plans
• Output: normalized 384-dimensional vectors

Evaluation is project-specific and does not prove traveler satisfaction. Structured constraints can improve duration/party/route matching but may also hide semantic-model weaknesses, so pure retrieval examples and failure cases are retained in the notebook.

Limitations

• English-only evaluation.
• Travel data contains synthetic fields and may be geographically imbalanced.
• Similarity is not evidence of safety, feasibility, current availability, or price.
• Users must verify live travel information with official sources.

License and attribution

Apache-2.0, inherited from the base model. Credit the original Sentence Transformers authors and Microsoft MiniLM work as documented in the upstream model card.