Text embeddings are a critical piece of many pipelines, from search, to RAG, to vector databases and more. Most embedding models are BERT/Transformer-based and typically have short context lengths (e.g., 512). That’s only about two pages of text, but documents can be very long – books, legal cases, TV screenplays, code repositories, etc can be tens of thousands of tokens long (or more). Here, we’re taking a first step towards developing long-context retrieval models .

We build on Monarch Mixer (M2), a recent model family developing attention- and MLP-free BERT models, which are enabling long-context BERT models. Today, we’re releasing a preview of a few models: long-context versions of M2-BERT up to 32K context length, as well as embedding versions fine-tuned for long-context retrieval. In this blog, you’ll hear about the changes to data mixture and loss function we needed to enable these new models.

We’re also releasing the V0 of a long-context retrieval benchmark called LoCo, that we’ve started to use to evaluate the long-context properties of various retrieval models.

Check out code here, and models up on HuggingFace here:

These models are also available on Together AI’s new embedding service – check it out here! Thanks to our collaborators at Together, we were able to beta-test these models at a MongoDB hackathon, and we already have early RAG integrations with LangChain and LlamaIndex.

We’ll have a full paper out next month – but for now, we wanted to release a preview to get early feedback from the community. Here are some particular calls to action for feedback if you’re interested in long-context retrieval:

If you have long-context retrieval tasks, we would love to hear how the M2-BERT retrieval models perform in the wild!

If you have public long-context retrieval tasks or datasets that you think would be good additions to LoCo, please let us know. We’ve only included a few retrieval tasks that have long documents, but we want to grow the benchmark to be more representative!