跳转至

Advanced RAG Techniques | 高级 RAG 技术

Naive RAG (Retrieve-Read) often fails with complex queries. Advanced RAG introduces optimization at pre-retrieval, retrieval, and post-retrieval stages.

朴素 RAG（检索-阅读）在处理复杂查询时往往表现不佳。高级 RAG 在检索前、检索中和检索后阶段引入了优化。

1. Pre-Retrieval Optimization | 检索前优化

Improving the user query to match the database better. 优化用户查询以更好地匹配数据库。

  • Query Expansion (查询扩展): Breaking a complex question into sub-questions.
    将复杂问题分解为子问题。
  • Query Transformation (查询转换): Rewriting the query to be more search-friendly.
    重写查询使其更适合检索。
  • User: "Compare revenue of Apple and Microsoft."
  • Transformed: "Apple revenue 2024", "Microsoft revenue 2024".

2. Retrieval Optimization | 检索优化

Finding the most relevant chunks. 找到最相关的切片。

  • Hybrid Search (混合检索): Combining Keyword Search (BM25) and Vector Search (Dense).
    结合关键词检索（BM25）和向量检索（稠密检索）。
  • Why: Vectors are good at semantics ("Apple" ~ "Fruit"), Keywords are good at exact match ("iPhone 15 Pro Max").
  • Metadata Filtering: Filtering by date, author, or category before searching.
    在搜索前按日期、作者或类别过滤。

3. Post-Retrieval Optimization | 检索后优化

Refining the results before feeding them to the LLM. 在将结果喂给 LLM 之前进行精炼。

  • Re-ranking (重排序): Using a specialized Cross-Encoder model (like BGE-Reranker) to score the relevance of retrieved chunks accurately.
    使用专门的交叉编码器模型（如 BGE-Reranker）精确评分检索到的切片的相关性。
  • Process: Retrieve top 100 -> Re-rank -> Keep top 5.
  • Context Compression: Removing irrelevant info from chunks to save tokens.
    从切片中移除无关信息以节省 Token。

4. GraphRAG (Knowledge Graphs) | 知识图谱 RAG

Instead of just vector similarity, GraphRAG uses Knowledge Graphs to understand relationships between entities. GraphRAG 不仅依赖向量相似度，还利用知识图谱来理解实体间的关系。

  • Use Case: "How does the CEO's decision affect the engineering team?" (Multi-hop reasoning).
    场景：“CEO 的决策如何影响工程团队？”（多跳推理）。
  • Mechanism: Traverses the graph edges to find connected concepts.
    机制：遍历图的边以找到连接的概念。
  • Vector DB: Milvus, Qdrant, Pinecone.
  • Framework: LlamaIndex (Best for RAG), LangChain.
  • Reranker: BAAI/bge-reranker-v2-m3.