Use case

RAG over documents

Embed the library, rerank the passages, then let a chat model answer from those passages only.

What matters

Retrieval quality usually matters more than which chat model writes the last paragraph. Long context is the fallback when you refuse to chunk.

accuracy · long-context · cost

#1Cohere: Embed 5 Fast
An embedding model for the index. Pair it with a reranker and a chat model for the answer.
#2NVIDIA: Nemotron 3.5 Embed 1B (free)
An embedding model for the index. Pair it with a reranker and a chat model for the answer.
#3LiquidAI: LFM2.5-Embedding-350M (free)
An embedding model for the index. Pair it with a reranker and a chat model for the answer.

Budget pick

NVIDIA: Nemotron 3.5 Embed 1B (free)

Open-weight pick

LiquidAI: LFM2.5-Embedding-350M (free)

Stuffing a whole wiki into a chat model skips the retriever and still misses the page.
Embeddings from one vendor and a reranker from another need a quick relevance check, not a new religion.

Answer only from the passages below. If they don't contain the answer, say so. Cite the passage ids.