chunking-strategies（分块策略）

Chunking Strategies for RAG

Optimize document splitting for retrieval accuracy and context preservation.

When to Use

• Designing a new RAG pipeline
• Retrieval quality is poor due to chunk boundaries
• Documents have mixed content types (code, tables, prose)
• Need to balance context window limits with retrieval precision

Chunking Methods

1. Fixed-Size Chunking

from langchain.text_splitter import CharacterTextSplitter

splitter = CharacterTextSplitter(
    chunk_size=1000,
    chunk_overlap=200,
    separator="\n"
)
chunks = splitter.split_text(document)

Best for: Homogeneous content, quick prototyping Avoid when: Documents have natural boundaries (sections, paragraphs)

2. Recursive Character Splitting

from langchain.text_splitter import RecursiveCharacterTextSplitter

splitter = RecursiveCharacterTextSplitter(
    chunk_size=1000,
    chunk_overlap=200,
    separators=["\n\n", "\n", ".", " ", ""]
)
chunks = splitter.split_documents(docs)

Best for: General-purpose text, maintains paragraph integrity Hierarchy: Tries larger separators first, falls back to smaller

3. Semantic Chunking

from langchain_experimental.text_splitter import SemanticChunker
from langchain_openai import OpenAIEmbeddings

splitter = SemanticChunker(
    embeddings=OpenAIEmbeddings(),
    breakpoint_threshold_type="percentile",
    breakpoint_threshold_amount=95
)
chunks = splitter.split_text(document)

Best for: When meaning matters more than size Trade-off: Slower, requires embedding calls

4. Document-Specific Chunking

Markdown

from langchain.text_splitter import MarkdownHeaderTextSplitter

headers = [
    ("#", "h1"),
    ("##", "h2"),
    ("###", "h3"),
]
splitter = MarkdownHeaderTextSplitter(headers_to_split_on=headers)
chunks = splitter.split_text(markdown_doc)

Code

from langchain.text_splitter import Language, RecursiveCharacterTextSplitter

splitter = RecursiveCharacterTextSplitter.from_language(
    language=Language.PYTHON,
    chunk_size=2000,
    chunk_overlap=200
)
chunks = splitter.split_documents(code_docs)

HTML

from langchain.text_splitter import HTMLHeaderTextSplitter

splitter = HTMLHeaderTextSplitter(
    headers_to_split_on=[("h1", "h1"), ("h2", "h2"), ("h3", "h3")]
)
chunks = splitter.split_text(html_doc)

Chunk Size Guidelines

Advanced: Parent-Child Chunking

from langchain.retrievers import ParentDocumentRetriever
from langchain.storage import InMemoryStore

# Small chunks for retrieval, large chunks for context
child_splitter = RecursiveCharacterTextSplitter(chunk_size=400)
parent_splitter = RecursiveCharacterTextSplitter(chunk_size=2000)

store = InMemoryStore()
retriever = ParentDocumentRetriever(
    vectorstore=vectorstore,
    docstore=store,
    child_splitter=child_splitter,
    parent_splitter=parent_splitter,
)

Why: Small chunks = precise retrieval, large chunks = better context

Metadata Enrichment

Always attach metadata to chunks:

for i, chunk in enumerate(chunks):
    chunk.metadata.update({
        "source": doc.metadata["source"],
        "chunk_index": i,
        "total_chunks": len(chunks),
        "doc_type": detect_doc_type(chunk.page_content),
        "has_code": bool(re.search(r'```', chunk.page_content)),
        "timestamp": datetime.now().isoformat()
    })

Evaluation Checklist

• Chunks don't break mid-sentence
• Code blocks stay intact
• Tables aren't split across chunks
• Headers stay with their content
• Overlap preserves context continuity
• Metadata enables filtering

Best Practices

1. Start with recursive splitting - works for 80% of cases
2. Test retrieval quality - not just chunk count
3. Use overlap - 10-20% prevents context loss at boundaries
4. Match chunk size to model - consider embedding model's optimal input
5. Preserve structure - use document-aware splitters when possible