Key takeaways

RAG with Chinese LLMs splits into embeddings (retrieval) and chat (generation). Pick embedding models that match your languages—CN-heavy corpora often favor Qwen/GLM paths; verify dimensions and rate limits on vendor consoles. This guide complements model selection (/en/articles/china-llm-rag-selection-guide) with setup steps, not another leaderboard.

Minimal RAG pipeline

Ingest → chunk (500–1k tokens with overlap) → embed → store (pgvector, Pinecone, etc.) → retrieve top-k → inject into chat prompt → generate with your primary China LLM. Log citations and failures. Optimize tokens (/en/articles/china-llm-optimization-guide-2026).

Vendor notes

Qwen/DashScope and others expose embedding endpoints—confirm names beside chat IDs (/en/articles/qwen-api-overseas-quickstart-2026). Do not assume OpenAI embedding parity—test CN/EN query-doc pairs. For compliance-sensitive docs, review /en/articles/china-llm-compliance-overseas-2026.

Next steps on Swift Horse

RAG selection /en/articles/china-llm-rag-selection-guide → SDK /en/articles/china-llm-openai-compatible-sdk-2026 → JSON /en/articles/china-llm-json-structured-output-2026.

FAQ

Which China LLM has the best embeddings?

No universal winner—evaluate recall@k on your bilingual corpus.

Can I use OpenAI embeddings with Chinese LLM chat?

Yes in hybrid stacks—ensure language alignment and compliance for stored vectors.

RAG vs long-context Kimi?

RAG scales cost for large corpora; long-context suits few huge docs—see /en/articles/kimi-api-overseas-quickstart-2026.

Is this official embedding documentation?

No—independent Swift Horse setup guide.