Text embeddings are dense vectors that represent the meaning of a sentence or paragraph. In RAG, the same model embeds both your document chunks and user queries so semantic search can find relevant passages. Choosing and tuning embeddings has a large impact on retrieval quality.
Look for strong performance on retrieval benchmarks, support for your languages, and a context length that matches your chunk size. Open-weight and API-hosted models both work; API models may be easier to start, while self-hosted models help with data residency.
Use one embedding model for indexing and querying. If you change models, re-embed the entire corpus. Mixing models breaks vector geometry and degrades search.
For non-English or mixed-language knowledge bases, pick a multilingual embedding model evaluated on your languages. For highly specialized jargon (medicine, law, engineering), test whether general embeddings suffice or domain-specific models improve recall.
Many pipelines L2-normalize vectors and use cosine similarity. Your vector database may expect a specific format—follow provider docs for best performance.
Embeddings sit between chunking and retrieval in how to build a RAG application. WeKnora wires embeddings, storage, and search together with LLM generation.