Benjamin Clavié

Stop Saying RAG Is Dead – Hamel’s Blog

2025-07-14T08:36:40Z

> Why the future of RAG lies in better retrieval, not bigger context windows.

Benjamin Clavié sur X : "What if a [MASK] was all you needed?..."

2025-02-11T00:25:23Z

A little pooling goes a long way for multi-vector representations – Answer.AI

2025-01-24T17:01:53Z

> Intuition: for documents focusing on a low number of topics, a lot of the tokens are likely to carry somewhat redundant semantic information, meaning keeping all of them is likely not useful.

Benjamin Clavié sur X : "Stella Embeddings: What's the big deal?..."

2025-01-13T18:42:18Z

> Training based on unsupervised distillation > The current dominant way of training retrieval models is via the use of a contrastive loss, with little-to-no knowledge distillation > (Stella's) training work within the embedding space, seeking to minimize the geometric distances... between the teachers' vectors and the student model (Stella)'s outputs. > > Stella models (and Jasper models) generalize amazingly well because of this.

[2412.13663] Smarter, Better, Faster, Longer: A Modern Bidirectional Encoder for Fast, Memory Efficient, and Long Context Finetuning and Inference

2024-12-21T22:45:32Z

Jeremy Howard sur X : "We trained 2 new models. Like BERT, but modern. ModernBERT. Not some hypey GenAI thing, but a proper workhorse model, for retrieval, classification, etc..."

2024-12-21T17:13:36Z

Benjamin Clavié sur X : "Introducing rerankers: a lightweight library to provide a unified way to use various reranking methods"

2024-03-16T10:28:38Z

bclavie/RAGatouille

2024-01-26T23:44:59Z

> RAGatouille's purpose is make it easy to use state-of-the-art methods in your RAG pipeline, without having to worry about the details or the years of literature! At the moment, RAGatouille focuses on making ColBERT simple to use. [Using ColBERT in-memory: Index-Free Encodings & Search](https://github.com/bclavie/RAGatouille/blob/0.0.5b1/examples/06-index_free_use.ipynb) ``` from ragatouille import RAGPretrainedModel RAG = RAGPretrainedModel.from_pretrained "colbert-ir/colbertv2. 0" ) # Your documents, a plain old list of chunked strings. documents = [...] # In-memory indexing supports metadata too! meta = ['attribute': ' really cool value'}...] # All the magic happens here RAG.encode documents, document_metadatas=meta) # Query your in-memory index RAG. search_encoded_docs(query = "A great question", k=3) # All further encode() calls add to the existing documents... RAG.encode(extra_documents, document_metadatas=extra_meta) # ... until you clear them RAG.clear_encoded ```