mistral-embed-2312

Embeddings can serve as features for nearest-neighbour or lightweight classifiers, but the model does not classify directly. Requires a downstream classification layer and labelled examples.

The 8192-token window accommodates full abstracts and long passages, so the model can support similarity-based retrieval of papers. It has no generation capability, so synthesis must be handled by a separate model.

Embedding-only model with no text generation capability. It can retrieve relevant passages but cannot produce answers.

No generative output capability, so it cannot hold a conversation. It could support retrieval of knowledge-base articles behind a separate chat model.

Usable only as the retrieval stage of a document Q&A pipeline. The model cannot produce answers, so a separate generative model is required for the complete use case.

The model returns vector representations and generates no text, so it cannot rewrite queries.

Bi-encoder embedding models score query and passage similarity independently and are not cross-encoder rerankers. Stronger results are expected from a dedicated reranking model.

Embedding model with no text generation. It cannot produce summaries.