chunkweaver

RAG chunker that respects document structure.

Install

pip install chunkweaver

Extras:

pip install chunkweaver[langchain]   # LangChain TextSplitter integration
pip install chunkweaver[llamaindex]  # LlamaIndex NodeParser integration
pip install chunkweaver[cli]         # CLI (included in base install)

Quick start

from chunkweaver import Chunker

# Minimal — just better defaults
chunker = Chunker(target_size=1024)
chunks = chunker.chunk(text)

# With structure-aware boundaries
from chunkweaver.presets import LEGAL_EU

chunker = Chunker(
    target_size=1024,
    overlap=2,
    overlap_unit="sentence",
    boundaries=LEGAL_EU,
)

chunks = chunker.chunk(text)

# With metadata
chunks = chunker.chunk_with_metadata(text)
for c in chunks:
    print(c.text)            # full chunk content
    print(c.start, c.end)    # character offsets in original
    print(c.boundary_type)   # "section" | "paragraph" | "sentence" | "word" | "keep_together"
    print(c.overlap_text)    # the overlap prefix (if any)
    print(c.content_text)    # text without overlap (for dedup)

Where chunkweaver fits

  PDF / DOCX / HTML              Your vector DB
        │                              ▲
        ▼                              │
  ┌───────────┐    ┌──────────────┐    │
  │ Extractor │───▶│ chunkweaver  │────┘
  │           │    │              │
  │ unstructured   │ boundaries   │  Embed + upsert
  │ marker-pdf     │ detectors    │  into Pinecone,
  │ docling        │ presets      │  Qdrant, Weaviate,
  │ pdfminer       │ overlap      │  ChromaDB, etc.
  └───────────┘    └──────────────┘

Your extractor turns files into text. Your vector DB stores embeddings. chunkweaver sits in the middle — splitting that text at structural boundaries so each chunk is a coherent unit of meaning, not an arbitrary slice of characters.

Why it matters

Standard chunkers, including LangChain's RecursiveCharacterTextSplitter, don't know that "Article 17" starts a new legal section, or that a table's header row belongs with its data. The result: chunks that straddle topic boundaries, producing blurry embeddings and incomplete retrievals.

The fix is cheap: surface markers (headings, article numbers, table rules) are reliable proxies for where topics change. Detecting them costs O(n) character comparisons — orders of magnitude less than computing semantic boundaries from embeddings.

Our LLM-as-judge benchmark on 11 documents across four domains and 58 queries:

Baseline	CW wins	Baseline wins	p-value
Naive 600-char	15	4	0.019
LangChain RCTS	11	4	0.119

See benchmarks/llm_judge/ for full results, methodology, and reproduction steps.

Features

• Zero dependencies — stdlib only, no LangChain/LlamaIndex tax
• Regex boundaries — you tell the chunker where sections start (^Article \d+, ^## , ^Item 1.)
• Hierarchical levels — CHAPTER > Section > Article > clause; split only as deep as needed
• Heuristic detectors — HeadingDetector, TableDetector discover structure from text patterns
• Annotation ingestion — accept pre-computed structure from any extractor
• Semantic overlap — sentences, not characters
• Full metadata — offsets, boundary types, hierarchy levels, overlap tracking
• Integrations — LangChain and LlamaIndex drop-ins

Presets

Built-in boundary patterns for common document types:

from chunkweaver.presets import (
    LEGAL_EU, LEGAL_US, RFC, MARKDOWN,
    CHAT, CLINICAL, FINANCIAL, FINANCIAL_TABLE,
    SEC_10K, FDA_LABEL, PLAIN,
)

Preset	Domain	Detects
LEGAL_EU	EU legislation	Article N, CHAPTER, SECTION, (1) recitals
LEGAL_US	US law / contracts	§ N, Section N, WHEREAS, 1.1 clauses
RFC	IETF RFCs	1. Intro, 3.1 Overview, Appendix A
MARKDOWN	Markdown	# headings, --- rules
CHAT	Chat logs	[14:30], ISO timestamps, speaker: turns
CLINICAL	Medical notes	HPI:, ASSESSMENT:, PLAN:, etc.
FINANCIAL	SEC filings	Item 1., PART I, NOTE 1, Schedule A
FINANCIAL_TABLE	Data tables	TABLE N, markdown/ASCII separators
SEC_10K	SEC annual reports	PART I–IV, Item N., ALL-CAPS sub-headings
FDA_LABEL	Drug labels	1 INDICATIONS, ## 2.1 Adult Dosage
PLAIN	Any	No boundaries — pure paragraph/sentence fallback

Combine presets freely:

boundaries = LEGAL_EU + [r"^TABLE\s+", r"^Annex\s+"]
boundaries = FINANCIAL + FINANCIAL_TABLE

Most presets have _LEVELED variants with hierarchical splits: LEGAL_EU_LEVELED, LEGAL_US_LEVELED, RFC_LEVELED, MARKDOWN_LEVELED, FINANCIAL_LEVELED, SEC_10K_LEVELED, FDA_LABEL_LEVELED. See docs/cookbook.md for details.

CLI

chunkweaver document.txt --size 1024 --overlap 2
chunkweaver legal_doc.txt --preset legal-eu --format json
chunkweaver file.txt --detect-boundaries --boundaries "^Article\s+\d+"
cat document.txt | chunkweaver --size 1024 --preset rfc
chunkweaver --recommend my_document.txt
chunkweaver --inspect my_document.txt

The --recommend flag analyzes a document and suggests the right config:

=== chunkweaver recommend ===

Document: 12,340 chars, 380 lines, 45 paragraphs

--- Preset matching ---
  legal-eu                8 hits <-- best

--- Detectors ---
  HeadingDetector: YES (12 headings found)
  TableDetector:   YES (2 tables found)

--- Python snippet ---
from chunkweaver import Chunker
from chunkweaver.presets import LEGAL_EU
from chunkweaver.detector_heading import HeadingDetector
from chunkweaver.detector_table import TableDetector

chunker = Chunker(
    target_size=1024,
    overlap=2,
    boundaries=LEGAL_EU,
    detectors=[HeadingDetector(), TableDetector()],
)

Integrations

LangChain

Drop-in replacement for RecursiveCharacterTextSplitter:

from chunkweaver.integrations.langchain import ChunkWeaverSplitter

splitter = ChunkWeaverSplitter(
    target_size=1024,
    overlap=2,
    boundaries=[r"^#{1,3}\s"],
)
docs = splitter.create_documents([text])

Requires: pip install chunkweaver[langchain]

LlamaIndex

Drop-in NodeParser for ingestion pipelines:

from chunkweaver.integrations.llamaindex import ChunkWeaverNodeParser
from chunkweaver.presets import LEGAL_EU

parser = ChunkWeaverNodeParser(
    target_size=1024,
    boundaries=LEGAL_EU,
    overlap=2,
)
nodes = parser.get_nodes_from_documents(documents)

Requires: pip install chunkweaver[llamaindex]

Heuristic detectors

For documents without clean section markers — SEC filings, scanned contracts, extracted PDFs — heuristic detectors discover structure from text patterns.

from chunkweaver import Chunker
from chunkweaver.detector_heading import HeadingDetector
from chunkweaver.detector_table import TableDetector

chunker = Chunker(
    target_size=1024,
    detectors=[HeadingDetector(), TableDetector()],
)

HeadingDetector scores lines on casing, length, whitespace context, and known prefixes. Works well on Title Case and ALL CAPS headings.

TableDetector identifies numeric data runs and marks them as keep-together regions. On SEC 10-K filings, keeps 80% of financial tables intact vs. 21% without it.

Custom detectors: subclass BoundaryDetector and return SplitPoint / KeepTogetherRegion. See examples/ml-detectors/ for scikit-learn examples.

Documentation

• Cookbook — domain recipes (clinical, FDA, SEC, financial, legal, chat, CJK), hierarchical boundaries, annotation ingestion, vector DB integration, tuning tips
• API Reference — full parameter tables, Chunk attributes, algorithm details, architecture
• FAQ — "What if I need to..." for common questions
• Benchmark — LLM-as-judge methodology, reproduction steps, raw results

Ecosystem

Part of a RAG tools suite for retrieval quality:

Tool	Role	What it measures
chunkweaver	Ingestion	Structure-aware chunking — controls what text enters the prompt
ragtune	Evaluation	Retrieval metrics (Recall@K, MRR, bootstrap CI) — measures how well your pipeline retrieves
ragprobe	Pre-deployment	Domain difficulty analysis — predicts how hard retrieval will be before you build

chunkweaver --recommend my_doc.txt              # what config to use
chunkweaver my_doc.txt --preset legal-eu \
    --export-dir ./chunks/                      # chunk → one .txt per chunk
ragtune ingest ./chunks/ --pre-chunked          # embed + store
ragtune simulate --queries golden.json          # measure retrieval
ragprobe analyze --corpus ./docs                # how hard is this domain

--export-dir writes ragtune-compatible files directly — no glue script needed. Use --format json with --recommend or --inspect for CI-friendly output.

Known limitations

• Sentence detection defaults to a simple regex ([.!?]\s+(?=[A-Z"(])). Abbreviations like "Dr. Smith" may cause false splits. For non-English or informal text, pass sentence_pattern — built-in alternatives: SENTENCE_END_CJK, SENTENCE_END_PERMISSIVE. See the cookbook for Cyrillic, Spanish, and other script-specific guidance.
• Boundaries are line-level regex matches — they won't detect inline structural markers.
• No tokenizer awareness — target_size is in characters, not tokens. For token budgets, estimate tokens ≈ chars / 4.

Author

Oleksii Alexapolsky (𝕏) — building retrieval quality tools: chunkweaver, ragtune, ragprobe.

License

MIT