contriever

Use cases

• Dense retrieval in domains lacking labeled query-document pairs
• Zero-shot dense retrieval baseline comparison
• Unsupervised passage retrieval for domain-specific corpora
• Initial retrieval stage in RAG pipelines where fine-tuning data is scarce
• Research into unsupervised vs. supervised dense retrieval tradeoffs

Pros

• Unsupervised training enables retrieval without any labeled data
• BERT backbone with standard HuggingFace transformers integration
• Publicly available Apache-adjacent weights for research
• Strong baseline for evaluating dense retrieval without supervision

Cons

• No pipeline_tag; requires manual transformers integration for inference
• Outperformed by supervised models (BGE, E5, nomic-embed) on standard benchmarks
• No instruction tuning or asymmetric query-passage training
• Domain-specific retrieval often requires fine-tuning despite unsupervised pretraining
• Less maintained than BAAI BGE and similar production-ready embedding models

Tags