Downloadanswerai-colbert-small-v1 is a 33M parameter multi-vector retrieval model developed by Answer.AI, using the JaColBERTv2.5 training recipe. Despite its small size (MiniLM scale), it surpasses many larger models—including e5-large-v2 and bge-base-en-v1.5—on standard information retrieval benchmarks. It is optimized for retrieval-augmented generation (RAG), reranking, and vector search, compatible with ColBERT, RAGatouille, and Rerankers libraries. The model achieves top performance in tasks like HotpotQA, TRECCOVID, and NQ, demonstrating strong zero-shot generalization. It is especially suited for efficient retrieval in low-resource environments or latency-sensitive applications. Pretrained and open-sourced under the Apache 2.0 license, the model supports ONNX and Safetensors for deployment flexibility. Its performance positions it as a practical solution for high-accuracy RAG pipelines without the compute overhead of large language models.
Features
- 33M parameter ColBERT model trained with JaColBERTv2.5
- Outperforms e5-large-v2 and bge-base on several BEIR tasks
- Compatible with ColBERT, RAGatouille, and Rerankers libraries
- Designed for multi-vector dense retrieval and reranking
- Supports ONNX and Safetensors formats for efficient deployment
- Provides fast and accurate retrieval with minimal memory footprint
- Ideal for RAG applications and passage reranking
- Open-source under Apache 2.0 license
Project Samples
