RAG准确率从61%到87%：chunk/embedding/rerank全链路调优实录

- Python 3.10.12
- langchain 0.1.0（只用它的Document抽象，不玩高级链）
- sentence-transformers 2.5.1（bge模型加载）
- FAISS CPU版 1.8.0.post1
- OpenAI GPT-3.5-turbo-0125（生成用，temperature=0）
- 文本切分：langchain.text_splitter + 自定义MarkdownHeaderSplitter
- Embedding候选：text-embedding-ada-002（维度1536） vs bge-large-zh-v1.5（维度1024）
- Rerank模型：bge-reranker-base（本地，约1.1GB显存）

from langchain.text_splitter import TextSplitter
import re

class InsuranceMarkdownSplitter(TextSplitter):
    def __init__(self, max_chunk_size=512, min_chunk_size=128):
        super().__init__()
        self.max_chunk_size = max_chunk_size
        self.min_chunk_size = min_chunk_size

    def split_text(self, text: str) -> list[str]:
        # 第一步：按二级标题切分（保留标题文本在chunk内）
        sections = re.split(r'(?m)^(##\s.*)$', text)
        chunks = []
        for i in range(1, len(sections), 2):
            header = sections[i]
            body = sections[i+1] if i+1  self.max_chunk_size and current_chunk:
                        if len(current_chunk) >= self.min_chunk_size:
                            chunks.append(current_chunk.strip())
                        current_chunk = ""
                    current_chunk += sent
                if current_chunk and len(current_chunk) >= self.min_chunk_size:
                    chunks.append(current_chunk.strip())
        return chunks

from sentence_transformers import SentenceTransformer
import faiss
import numpy as np

# 加载bge模型（本地）
model = SentenceTransformer('BAAI/bge-large-zh-v1.5', device='cuda')
# 注意：bge在encode时如果不加query指令，检索效果会下降5%左右
query_instruction = "为这个句子生成表示以用于检索相关文章："

def build_index(all_chunks: list[str]):
    # 为所有chunk生成向量，bge不需要加指令，只有query需要
    chunk_vectors = model.encode(all_chunks, batch_size=32, 
                                 normalize_embeddings=True, device='cuda')
    index = faiss.IndexFlatIP(1024)  # 向量维度变成了1024
    index.add(chunk_vectors)
    faiss.write_index(index, 'insurance_bge.index')
    return index

def search(query: str, index, top_k=10):
    # 关键：查询端要加指令前缀
    q_vector = model.encode([query_instruction + query], 
                            normalize_embeddings=True, device='cuda')
    scores, indices = index.search(q_vector, top_k)
    return scores[0], indices[0]

from BCEmbedding import RerankerModel

reranker = RerankerModel(model_name_or_path="BAAI/bge-reranker-base", device='cuda')

def rerank_and_generate(query: str, top_k=3):
    # 1. 粗排取Top10
    _, indices = search(query, index, top_k=10)
    candidates = [all_chunks[i] for i in indices]

    # 2. 精排：对(query, chunk)对打分，返回相关性分数
    pairs = [[query, doc] for doc in candidates]
    scores = reranker.compute_score(pairs)

    # 3. 按分数降序，取Top3
    sorted_pairs = sorted(zip(candidates, scores), 
                          key=lambda x: x[1], reverse=True)
    top3 = [doc for doc, score in sorted_pairs[:top_k]]

    # 4. 拼Prompt给LLM
    context = "\n\n---\n\n".join(top3)
    prompt = f"基于以下保险条款片段回答问题，不要编造：\n\n{context}\n\n问题：{query}"
    # ... 调用GPT-3.5