[Official FAQ] Share your experience or follow-up questions below.

The embedding model decides whether the right passages are found, so it often matters more than the chat model for RAG quality.

What to look at

• Language support: if your documents are not only English, choose a model that explicitly supports multiple languages. English-focused models often do poorly on other languages.
• Candidates people commonly try: the BGE family (for example bge-m3, which is multilingual and handles longer inputs), the Qwen embedding models, multilingual-e5, and nomic-embed-text (often used for English-heavy content). New versions appear quickly, so check each model card and test.
• Max input length: text longer than the limit is truncated, so keep chunks within it.
• Size and speed: larger models can be more accurate but make indexing slower.
• Vector dimensions: higher dimensions cost more storage and search time.

Important usage rules

• Use the same embedding model for indexing and querying. If you change it, rebuild the index.
• Some models expect different instruction prefixes for queries and documents. Read the model card, because skipping this can hurt results.
• Ollama and LM Studio can both serve embedding models through their APIs.

How to decide: prepare 20–30 realistic questions, mark which passages should be found, and compare hit rates between two or three embedding models. Adding a reranker on top often gives a further boost.

Which one are you using, and in which languages?