Embedding and Reranking Models — References

Benchmarks, model families, and the retrieval architectures behind them.

Benchmarks

• MTEB: Massive Text Embedding Benchmark (Muennighoff et al., 2022), arXiv:2210.07316 — the benchmark and its task taxonomy; read the per-task tables rather than the average.
• MTEB leaderboard — current standings, useful for shortlisting and unsafe as a selection criterion on its own.
• BEIR: A Heterogeneous Benchmark for Zero-shot Evaluation of Information Retrieval Models (Thakur et al., 2021), arXiv:2104.08663 — zero-shot retrieval across domains, and the evidence that domain transfer is uneven.

Architectures

• Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks (Reimers & Gurevych, 2019), arXiv:1908.10084 — the bi-encoder construction, and why a cross-encoder cannot be used for first-stage retrieval.
• ColBERT: Efficient and Effective Passage Search via Contextualized Late Interaction over BERT (Khattab & Zaharia, 2020), arXiv:2004.12832 — late interaction as a middle point between bi-encoders and cross-encoders.
• Matryoshka Representation Learning (Kusupati et al., 2022), arXiv:2205.13147 — the training method behind embeddings whose leading dimensions can be truncated without retraining.
• Reciprocal Rank Fusion outperforms Condorcet and individual Rank Learning Methods (Cormack, Clarke & Buettcher, SIGIR 2009) — the fusion rule used for hybrid retrieval. Cited by name; it is a conference paper without a stable open identifier.

Implementations

• Sentence-Transformers documentation — bi-encoder and cross-encoder usage, pooling strategies, and normalisation, which is where most reproducibility mismatches originate.
• Text Embeddings Inference — a server for open-weight embedding and reranking models.

Individual model cards on the Hugging Face Hub are the authority on each family's required instruction prefixes and maximum sequence length; they are not linked individually here because the correct card is the one for the exact revision you deploy.