RAG召回率从61%到89%：chunk重切、embedding换代与rerank落地的完整记录

Python 3.10.12
LlamaIndex 0.10.43
langchain 0.2.9
sentence-transformers 2.7.0
FAISS 1.7.4
torch 2.2.1+cu118

from llama_index.core.node_parser import HierarchicalNodeParser, SentenceSplitter
from llama_index.core.schema import Document

# 核心：按标题层级构建父节点，子节点按句子切分
node_parser = HierarchicalNodeParser.from_defaults(
    chunk_sizes=[512, 256, 128],  # 从大到小，形成层级
    chunk_overlap=32,
    include_prev_next_rel=True  # 保留前后节点关系
)

doc = Document(text=raw_text, metadata={"source": file_name})
nodes = node_parser.get_nodes_from_documents([doc])

# 打印切割后的chunk示例，验证是否保留了标题路径
for node in nodes[:3]:
    print(f"Node ID: {node.node_id}")
    print(f"Metadata: {node.metadata}")
    print(f"Text snippet: {node.get_content()[:50]}...")

def add_heading_path(nodes):
    for node in nodes:
        if node.relationships:
            parent = node.relationships.get("parent")
            if parent:
                node.metadata["heading_path"] = parent.metadata.get("heading_path", "") + " > " + parent.metadata.get("heading", "")
    return nodes

nodes = add_heading_path(nodes)

from llama_index.core import Settings
from llama_index.embeddings.huggingface import HuggingFaceEmbedding

# 切换到BGE-M3，注意设置query和passage的指令
embed_model = HuggingFaceEmbedding(
    model_name="BAAI/bge-m3",
    device="cuda:0",
    embed_batch_size=32,
    query_instruction="为这个句子生成表示以用于检索相关文章：",
    text_instruction="为这个句子生成表示："
)
Settings.embed_model = embed_model

# 重建向量索引（必须重新embedding所有文档）
from llama_index.core import VectorStoreIndex, StorageContext
from llama_index.vector_stores.faiss import FaissVectorStore
import faiss

dim = 1024  # BGE-M3输出维度
faiss_index = faiss.IndexFlatIP(dim)
vector_store = FaissVectorStore(faiss_index=faiss_index)
storage_context = StorageContext.from_defaults(vector_store=vector_store)

index = VectorStoreIndex.from_documents(
    documents, 
    storage_context=storage_context,
    embed_model=embed_model
)

from llama_index.core.postprocessor import SentenceTransformerRerank

# 初始化rerank模型，注意用专门的bge-reranker，不是embedding模型
reranker = SentenceTransformerRerank(
    model="BAAI/bge-reranker-v2-m3",
    top_n=3,  # 保留前3个
    keep_retrieval_score=True
)

# 查询时使用
from llama_index.core import QueryBundle

query_bundle = QueryBundle(query_str="如何修改付款审批人的权限？")
retrieved_nodes = index.as_retriever(similarity_top_k=5).retrieve(query_bundle)
reranked_nodes = reranker.postprocess_nodes(retrieved_nodes, query_bundle)

for node in reranked_nodes:
    print(f"Score: {node.score:.4f} | Text: {node.get_content()[:60]}...")

# 优化前：直接Top-5进行rerank，浪费了前面检索的潜力
# 优化后：先取Top-20，再rerank出Top-3
retriever = index.as_retriever(similarity_top_k=20)
reranker = SentenceTransformerRerank(model="BAAI/bge-reranker-v2-m3", top_n=3)