Building RAG Pipeline with Chinese AI Models

Jul 27, 2026

title: "Building a RAG Pipeline with Chinese AI Models"

published: true

tags: rag, chinese-ai, langchain, embedding

canonical_url: https://aiwave.live/blog/rag-pipeline-chinese-ai-models

description: "Build a RAG pipeline using Chinese AI models like DeepSeek and GLM. Covers embedding selection, chunking strategies, and a complete Python implementation without LangChain."

RAG (Retrieval-Augmented Generation) combines document retrieval with LLM generation to answer questions grounded in your data. Most RAG tutorials use GPT-4. Chinese AI models work just as well — and cost a fraction of the price.

Available through AIWave with OpenAI-compatible APIs.

Architecture

Documents → Chunking → Embedding → Vector Store → Query Embedding → Similarity Search → LLM → Answer

Embedding Model Selection

For Chinese text, embedding model choice matters:

Model	Dimensions	Chinese	Cost	Source
text-embedding-3-small	1536	OK	$0.02/1M tokens	OpenAI
bge-m3	1024	Excellent	Free (local)	BAAI
bge-large-zh-v1.5	1024	Excellent	Free (local)	BAAI

For Chinese-heavy content, BGE models (from BAAI) outperform OpenAI's embeddings. bge-m3 supports multilingual text and is small enough to run locally.

Chunking Strategy

Chinese text chunking differs from English:

No word-level tokenization: Chinese doesn't have spaces between words. Chunk by character count (500-1000 chars) rather than word count

Paragraph boundaries: Prefer splitting at paragraph breaks (\n\n) to maintain semantic coherence

Overlap: 100-200 character overlap between chunks to avoid losing context at boundaries

def chunk_text(text, chunk_size=800, overlap=200):
    chunks = []
    start = 0
    while start < len(text):
        end = start + chunk_size
        chunks.append(text[start:end])
        start = end - overlap
    return chunks

Vector Store

Use FAISS (local, free) or a hosted solution:

import faiss
import numpy as np
from openai import OpenAI

client = OpenAI(api_key="***", base_url="https://aiwave.live/v1")

embeddings = []
for chunk in chunks:
    resp = client.embeddings.create(
        model="text-embedding-3-small",
        input=chunk
    )
    embeddings.append(resp.data[0].embedding)

index = faiss.IndexFlatIP(len(embeddings[0]))
index.add(np.array(embeddings).astype('float32'))

Query and Generation

def query_rag(question, index, chunks, top_k=3):
    resp = client.embeddings.create(
        model="text-embedding-3-small",
        input=question
    )
    query_vec = np.array([resp.data[0].embedding]).astype('float32')
    distances, indices = index.search(query_vec, top_k)
    context = "\n\n".join([chunks[i] for i in indices[0]])
    
    response = client.chat.completions.create(
        model="deepseek-chat",
        messages=[
            {"role": "system", "content": "Answer based on the provided context."},
            {"role": "user", "content": f"Context:\n{context}\n\nQuestion: {question}"}
        ]
    )
    return response.choices[0].message.content

Cost

Using DeepSeek V4 Flash ($0.14/$0.28) for generation and text-embedding-3-small for retrieval:

100 documents × 10 chunks × 1536-dim embedding: ~$0.01

1,000 RAG queries × 4K context average: ~$5/month

Compare to GPT-4o: ~$60/month for the same workload.

All components are available through AIWave with a single API key. The $1 free credit covers your first 5,000+ RAG queries. Replace LangChain with 100 lines of standard Python — simpler, faster, no framework lock-in.