DocSlicer

Lightning-fast (31 pages/sec), deterministic document parser and chunker for business documents. No LLM calls or heavy ML models.

DocSlicer turns PDFs, Word documents, HTML pages, and PowerPoint files into clean chunks, structured blocks, tables, charts, markdown and a navigable heading hierarchy.

Top score on BizDocBench (0.88 overall vs 0.70 for the next-best tool). 0.80 table accuracy, 0.98 content faithfulness, 0.85 heading recognition and hierarchy preservation, and 0.76 RAG retrieval performance.

Two ways to use it:

• As a Python library — classic RAG. The layout-aware chunker gives you clean, non-overlapping chunks, each carrying its full heading breadcrumb, ready to embed. Jump to the API ↓
• As an MCP server — vectorless RAG. For when you want an answer out of a document right now. Claude, Cursor, or VS Code pulls the outline, picks the section it needs, and reads only that — no embedding, and no 200-page document in the context window. Jump to setup ↓

Quick start

import docslicer

def main():
    result = docslicer.parse_document("annual_report.pdf")

    # Inspect the outline first
    result.hierarchy.to_outline()
    # - PART I — FINANCIAL INFORMATION
    #   - Item 1. Financial Statements
    #     - Notes to Condensed Consolidated Financial Statements
    #       - Note 4 – Financial Instruments
    #         - Derivative Instruments and Hedging
    #           - Foreign Exchange Rate Risk
    #           - Interest Rate Risk
    #         - Accounts Receivable
    #           - Trade Receivables
    #   - Item 2. Management's Discussion and Analysis
    #     - Liquidity and Capital Resources
    # - PART II — OTHER INFORMATION
    #   ...

    # Pull only the chunks you need
    risk_section = result.find_heading("Risk Factors")[0]
    chunks = result.chunks_under(risk_section)

    # Tables come back structured, not as flat text
    for table in result.tables_under(risk_section):
        print(table.markdown)

if __name__ == "__main__":
    main()

Features

• No LLM, VLM, or ML models — fully deterministic; no model weights to download, no GPU required, no cold-start delay
• Lightweight — ~630 KB wheel with no heavy ML dependencies
• Agentic-friendly — reduces token spend on long documents: have the agent inspect the outline first, then pull only the relevant chunks into context instead of feeding a 500-page document verbatim; well-suited for legal texts, technical SOPs, financial filings, and compliance documents
• Deep hierarchy extraction — works for both numbered (1., 1.2., 1.2.3) and free-form headings; uses font size, bold weight, and document structure — not inference; handles re-entry after exhibit breaks and repeated navigation headings across pages
• Structure-aware chunking — splits at heading and paragraph boundaries, preserving semantic coherence
• Zero character overlap — chunks are non-overlapping by default; no duplicated tokens in your context window
• Unified result object — chunks, blocks, tables, charts, metadata, and hierarchy in one place
• Structured tables — tables come back as cells, not flat text; export as Markdown, JSONL, or melted format
• Multiple export formats — CSV, Markdown, JSONL, Parquet, JSON, plain text, and DataFrames
• Reading order preserved — including multi-column PDF layouts
• Supports pdf, docx, pptx, and html — including JS-rendered pages via Playwright
• Robust URL fetching — always renders pages in a real browser, handling cookie banners and bot protection out of the box; also preserves styling signals like boldness that raw HTML omits, producing sharper heading detection and chunk quality
• OCR fallback — auto-detects scanned pages and falls back to Tesseract when the extra is installed

Benchmarks

Measured with BizDocBench — an open benchmark for multi-format business document parsing. All scores are 0–1 (higher is better); pages_per_sec_aggregate is throughput across the full corpus.

Install

pip install docslicer

The core install is dependency-light. Optional features are available as extras:

pip install 'docslicer[html]'    # HTML / URL parsing via Playwright
playwright install chromium       # one-time browser install (Chromium only)

pip install 'docslicer[ocr]'     # scanned PDF support via Tesseract + OpenCV
# The tesserocr wheel bundles libtesseract but NOT the language models,
# so install the Tesseract engine to provide them (docslicer auto-detects the path):
# Linux:  apt install tesseract-ocr
# macOS:  brew install tesseract

pip install 'docslicer[mcp]'     # MCP server for LLM clients (Claude, Cursor, …)
pip install 'docslicer[llm]'     # exact token counts via tiktoken (exact_tokens=True)
pip install 'docslicer[crypto]'  # password-protected Office files (msoffcrypto-tool)
pip install 'docslicer[parquet]' # Parquet export support

Extras can be combined: pip install 'docslicer[html,ocr,llm]'.

Requires Python 3.10+

What you get back (ParseResult)

parse_document returns a ParseResult:

result.chunks      # list[Chunk]   — heading-aware text chunks, ready for embedding
result.blocks      # list[Block]   — paragraph/heading/table blocks before chunking
result.tables      # list[Table]   — structured tables with cells, spans, and markdown
result.charts      # list[Chart]   — charts as extracted data points (docx/pptx)
result.metadata    # DocumentMetadata — title, author, language, page count, OCR flag
result.hierarchy   # HierarchyTree — navigable tree of all headings

Each Chunk carries:

chunk.text          # str   — chunk text
chunk.path          # list  — full heading breadcrumb from root to nearest heading
chunk.heading       # str   — nearest heading above this chunk
chunk.section       # str   — body | toc | exhibit | header | footer | coverpage | …
chunk.page_number   # int   — 1-based physical page
chunk.page_label    # str   — "A-6", "iv", "F-3" — as printed on the page
chunk.table_ids     # list  — IDs of tables referenced in this chunk
chunk.chart_ids     # list  — IDs of charts referenced in this chunk (docx/pptx)
chunk.link_url      # list  — URLs found in this chunk
chunk.bbox          # BBox  — bounding box (PDF only)

Every chunk carries its full heading breadcrumb, no matter how deeply nested. For example, a paragraph six levels deep in a financial filing:

chunk.path == [
    "# PART I — FINANCIAL INFORMATION",
    "## Item 1. Financial Statements",
    "### Notes to Condensed Consolidated Financial Statements (Unaudited)",
    "#### Note 4 – Financial Instruments",
    "##### Accounts Receivable",
    "###### Trade Receivables",
]

This lets downstream code filter or group chunks by any level of the hierarchy without re-parsing the document.

Supported formats

Not supported: .doc, .ppt (legacy Office formats), .xlsx.

Parsing

parse_document auto-detects the format from the file extension or magic bytes. Pass a file path, URL, raw bytes, or a file-like object:

result = docslicer.parse_document("contract.docx")
result = docslicer.parse_document("report.pdf")
result = docslicer.parse_document("https://www.sec.gov/Archives/edgar/data/.../10-K.htm")
result = docslicer.parse_document(file_bytes)

Parsing & content options

parse_document (and the format-specific functions) accept options that control what gets parsed and how, before it's chunked. Format-specific toggles are accepted everywhere for a uniform API but only take effect for the relevant format.

result = docslicer.parse_document(
    "contract.docx",
    password="admin123",           # decrypt password-protected files; .docx and .pptx needs [crypto] extra
    max_workers=4,                 # process-pool width for PDF extraction/OCR (default: auto by CPU cores)
    include_headers_footers=True,  # docx: include header/footer content (default False)
    include_footnotes=True,        # docx: include footnotes (default True)
    include_comments=True,         # docx: include review comments (default False)
    include_speaker_notes=True,    # pptx: include slide speaker notes (default True)
    use_browser=True,              # html/URL: render in a real browser (default True)
)

Chunking options

result = docslicer.parse_document(
    "report.pdf",
    max_chunk_size=2000,          # hard cap, default 3200
    optimal_chunk_size=800,       # target size, default 1500
    min_chunk_size=400,           # soft floor, default 700
    chunking=False,               # skip chunking, return blocks only (faster)
    merge_small_chunks=True,      # merge chunks below min_chunk_size (default True)
    table_representation="jsonl", # "markdown" (default) | "jsonl" | "melted"
    exact_tokens=True,            # exact tiktoken (cl100k_base) counts; needs [llm] extra, else char/4 estimate
    extra_fields=["is_bold", "font_size", "font_name"],  # surface internal pipeline columns on each chunk/block via .extra
)

How merge_small_chunks works

Because DocSlicer is structure-aware, it initially produces one chunk per heading or paragraph boundary. For documents with many short sections this can yield a lot of small chunks. With merge_small_chunks=True (the default), sibling sections under the same parent heading are merged together until they reach min_chunk_size — but never across heading boundaries into a different parent.

For example, these five short sections all fall under ## Products and Services Performance:

### Mac          → "Mac net sales decreased …"            (~120 chars)
### iPad         → "iPad net sales increased …"           (~180 chars)
### Wearables    → "Wearables net sales decreased …"      (~130 chars)
### Services     → "Services net sales increased …"       (~160 chars)

Instead of four tiny chunks, they get merged into one coherent chunk that still carries the correct path for each paragraph. Set merge_small_chunks=False if you need one chunk per section regardless of size.

Table representation formats

table_representation controls how tables are serialised into chunk text. Given a financial table with multi-row column headers:

"markdown" (default) — preserves the original 2D layout:

|           | Three Months Ended    | Three Months Ended    |
|           | December 27, 2025     | December 28, 2024     |
|-----------|----------------------:|----------------------:|
| iPhone ®  |              $85,269  |              $69,138  |
| Mac ®     |               8,386   |               8,987   |
| iPad ®    |               8,595   |               8,088   |
| …         |                   …   |                   …   |

"melted" — one row per cell, headers joined with >. Good for sparse or pivot-style tables where individual cell retrieval matters:

iPhone ® | Three Months Ended > December 27, 2025 | $85,269
iPhone ® | Three Months Ended > December 28, 2024 | $69,138
Mac ® | Three Months Ended > December 27, 2025 | 8,386
Mac ® | Three Months Ended > December 28, 2024 | 8,987
iPad ® | Three Months Ended > December 27, 2025 | 8,595
iPad ® | Three Months Ended > December 28, 2024 | 8,088
…

"jsonl" — one JSON object per row, multi-row headers joined with _. Useful when chunks are fed into structured extraction or tool-use pipelines:

{"Metric": "iPhone ®", "Three Months Ended_December 27, 2025": "$85,269", "Three Months Ended_December 28, 2024": "$69,138"}
{"Metric": "Mac ®", "Three Months Ended_December 27, 2025": "8,386", "Three Months Ended_December 28, 2024": "8,987"}
{"Metric": "iPad ®", "Three Months Ended_December 27, 2025": "8,595", "Three Months Ended_December 28, 2024": "8,088"}
…

Batch processing

Point parse_all at a folder (or pass a list of paths/URLs). It yields (source, result) pairs, and a file that fails to parse yields the Exception instead of aborting the batch. Any parse_document keyword — chunk sizes, include_*, etc. — is forwarded per document.

for path, result in docslicer.parse_all("documents/", recursive=True, max_chunk_size=2000):
    if isinstance(result, Exception):
        print(f"Failed {path}: {result}")
    else:
        print(f"{path}: {len(result.chunks)} chunks")

Reuse config across documents

DocumentParser holds a fixed ParseConfig across many documents and keeps a single browser open across HTML/URL inputs (launched lazily on the first HTML parse), so a batch of URLs starts Chromium once instead of once per document. Use it as a context manager so that browser is always released:

from docslicer import DocumentParser, ParseConfig

config = ParseConfig(max_chunk_size=1500, optimal_chunk_size=600)

with DocumentParser(config) as parser:
    for path, result in parser.parse_all(paths):   # or parser.parse(path) for one
        ...

Two levels of parallelism

There are two independent knobs, and they compose:

• ParseConfig(max_workers=N) — within a single document: parallelizes PDF word extraction, cell building, and OCR across processes (default: auto, sized to CPU cores). Best when documents are large.
• DocumentParser(config, workers=N) — across documents: fans whole documents out over N worker processes, each with its own config and browser. Best when you have many documents. Results arrive in submission order (this path isn't lazy per-document).

Setting workers alone defaults each worker's max_workers to 1, so nested pools don't oversubscribe the machine; set both explicitly to run both levels at once. The workers path can't forward a browser session or on_stage callback across processes — leave workers unset when you need those.

Guard your entry point. DocSlicer uses a ProcessPoolExecutor whenever there's real CPU work to fan out — any PDF over ~50 pages, any scanned/OCR PDF of any length, and both parallelism knobs above. This is not opt-in: a plain docslicer.parse_document("big.pdf") triggers it too. On macOS and Windows, Python spawns workers by re-importing your script top to bottom, so a parse that runs at module level makes each worker re-run it and spawn again — raising RuntimeError: An attempt has been made to start a new process before the current process ... bootstrapping phase. Put your parsing code inside a function behind an if __name__ == "__main__": guard:

python

Copy

def main():
    with DocumentParser(config, workers=4) as parser:
        for path, result in parser.parse_all(paths):
            ...

if __name__ == "__main__":
    main()

Navigating the hierarchy

Most chunking libraries give you a flat list of text segments. DocSlicer also gives you a navigable tree of the document's heading structure, extracted deterministically from the document itself.

This is particularly useful for agents and retrieval pipelines working with long documents: rather than feeding the entire document into context, the agent can inspect the outline first to understand the structure, decide which sections are relevant, and then pull only those chunks — keeping token usage proportional to the task.

Inspect the outline

# Print the full heading tree
result.hierarchy.to_outline()

# Walk all top-level sections and see how much content each contains
for node in result.hierarchy.level(1):
    print(node.text, "→", len(result.chunks_under(node)), "chunks")

Drill into a section

.level(n) returns all headings at depth n. Pass a parent to scope it to a specific subtree — the typical pattern for an agent navigating a long document:

# All top-level headings
l1 = result.hierarchy.level(1)

# Pick one, then list its subsections
section = result.find_heading("Financial Statements")[0]
for node in result.hierarchy.level(2, parent=section):
    print(node.text, f"(p.{node.page_number})")

# Drill one level deeper
subsection = result.hierarchy.level(2, parent=section)[0]
for node in result.hierarchy.level(3, parent=subsection):
    print(node.text)

Retrieve content under a heading

find_heading matches any node whose text contains the search term (case-insensitive). All retrieval methods recurse into subsections by default.

node = result.find_heading("Financial Instruments")[0]

chunks = result.chunks_under(node)              # text chunks, ready for embedding or prompting
chunks = result.chunks_under(node, recursive=False)  # direct heading only, no subsections
tables = result.tables_under(node)              # structured tables in this section
charts = result.charts_under(node)              # charts (with extracted data points) in this section
blocks = result.blocks_under(node)              # raw paragraph/heading blocks

Navigate by page

result.chunks_by_page(14)        # by page number
result.chunks_by_page("F-3")     # by printed page label
result.blocks_by_page(14)
result.tables_by_page(14)
result.charts_by_page(14)

Parse once, navigate many times

A parsed result is plain data, so you can persist it and reload it later. When an agent asks many questions about the same document, there's no need to parse it again on every question:

from pathlib import Path
import docslicer

cache = Path("annual_report.json")

if cache.exists():
    result = docslicer.ParseResult.load(cache)
else:
    result = docslicer.parse_document("annual_report.pdf")
    result.save(cache)

A reloaded result supports the full API — hierarchy, find_heading, chunks_under, tables — so a long-running agent session or document server can keep documents open across requests without re-parsing.

Export

save() decides what to write from the path you give it.

# Save the whole result and reload it later — keeps the heading hierarchy
result.save("result.json")                     # same output as result.to_json()
result = docslicer.ParseResult.load("result.json")

# A single collection, in the format you name
result.save("chunks.csv")
result.save("charts.jsonl")       # stems: chunks | blocks | tables | charts | metadata
result.export_chunks_jsonl("chunks.jsonl")

# One file per collection
result.save("output/")
# → output/chunks.parquet, blocks.parquet, tables.parquet, metadata.json
#   (+ charts.parquet when the document has charts)
#   Falls back to .csv unless the [parquet] extra is installed.

# Render as Markdown or plain text
md = result.export_to_markdown(include_tables=True)
txt = result.export_to_text()

# DataFrames
df = result.chunks_df()

Only result.json round-trips — the collection and directory forms write flat rows without the heading hierarchy, so ParseResult.load() can't read them back.

Debug mode

result = docslicer.parse_document("report.pdf", debug=True)

# result.pipeline_steps is an ordered dict of step name → DataFrame
for name, df in result.pipeline_steps.items():
    print(name, df.shape)
    df.to_csv(f"debug/{name}.csv", index=False)

# PDF steps:        words → shapes → cells → lines → table_cells → blocks → chunks
# DOCX/PPTX steps:  runs → chart_points → paragraphs → lines → table_cells → blocks → chunks

OCR

parse_document automatically detects scanned pages and falls back to OCR when the [ocr] extra is installed. No configuration needed — result.metadata.has_ocr tells you whether OCR was used.

pip install 'docslicer[ocr]'
# tesserocr binds libtesseract directly, so install the Tesseract dev libraries first:
# Linux:  apt install tesseract-ocr libtesseract-dev libleptonica-dev pkg-config
# macOS:  brew install tesseract leptonica

MCP server

DocSlicer ships an MCP server, so LLM clients (Claude Desktop, Claude Code, Cursor, …) can parse and read documents directly.

pip install 'docslicer[mcp]'
docslicer-mcp                    # stdio — what desktop clients launch
docslicer-mcp --transport http --port 8000

Claude Desktop / Cowork — one-click install

Download docslicer-X.Y.Z.mcpb from the latest release and double-click it, or drag it onto the Claude Desktop window. You pick the folder DocSlicer is allowed to read and write during install; no config file, and no Python of your own — uv provisions the interpreter.

Other clients

Every client below launches the server over stdio. uvx needs nothing installed ahead of time:

{
  "mcpServers": {
    "docslicer": {
      "command": "uvx",
      "args": ["--from", "docslicer[mcp]", "docslicer-mcp"],
      "env": { "DOCSLICER_MCP_ROOT": "/Users/you/Documents" }
    }
  }
}

If you'd rather install it once and skip the resolve on every launch, use pip install 'docslicer[mcp]' (or uv tool install) and set "command": "docslicer-mcp" with no args.

On GUI-launched clients, prefer the .mcpb. An app started from the dock does not inherit your shell PATH — on macOS that excludes /opt/homebrew/bin — so a bare uvx or docslicer-mcp can work in a terminal and fail when the client spawns it. Use an absolute path (which uvx) if you hit this. The extension sidesteps it entirely.

How it works

A parsed document is far larger than a model's context window, so the server never returns one in a single call. parse registers the document and hands back a doc_id handle plus a heading outline. Every other tool takes that handle and returns a bounded slice — the model pulls in only what it needs.

Every outline line carries what reading it would cost:

- Financial statements  ~48k
  - Note 14 — Segment reporting  ~900
  - Note 15 — Income taxes  ~2.1k

That figure is the same estimate read reports back, so a budget made from the outline holds when it is spent. Sizes are cumulative — a parent never costs less than the children beneath it — which is what makes "descend or just read it" a decision the model can make before spending the context rather than after.

read takes heading text exactly as the outline prints it. Where a heading appears twice, prefixing any ancestor disambiguates it ("Notes > Revenue"); the full chain is never required. Returned text is interleaved with [Page X] markers using the document's own page labels (S-23, iv), so a quotation can be cited to the page it actually came from rather than to wherever its section began.

search is the fallback for when the outline does not settle the question — headings that name nothing useful (Note 14, Item 7A), or a figure buried in a table no heading mentions. It combines a whole-word literal match with BM25 over the chunks, and returns places, not answers: each hit is a heading to pass to read. Query terms that appear nowhere in the document are reported back, so a query that scored well on one rare word can be recognised as the bad query it was.

to_markdown is the escape hatch for when the user wants the document itself rather than an answer drawn from it. It writes to disk and returns a path, so nothing enters the model's context and document size stops mattering.

Parsed results are cached on disk, so re-parsing the same file with the same options is free. The cache key includes the file's size and mtime — edit the document and the next parse re-parses it automatically.

Configuration

Set DOCSLICER_MCP_ROOT when exposing the server to anything but yourself — without it, any readable path on the machine is parseable, and to_markdown can write anywhere the server process can.

Command line

docslicer parses one document to JSON on stdout — for a quick look at a file, or to pipe into jq.

docslicer report.pdf                   # chunks as JSON
docslicer report.pdf -o chunks.json    # write to a file
docslicer report.pdf --no-chunking     # blocks instead of chunks

It takes the same parsing and chunking options as parse_document; run docslicer --help for the full list.

Format-specific functions

If you know the format upfront and want explicit failure on unexpected input, use the format-specific variants. They accept the same arguments as parse_document:

docslicer.parse_pdf("report.pdf")
docslicer.parse_docx("contract.docx")
docslicer.parse_pptx("deck.pptx")
docslicer.parse_html("filing.html")

Privacy Policy

Full policy: https://docslicer.ai/privacy

What is collected. Nothing. DocSlicer has no telemetry, analytics, crash reporting, or usage tracking, and requires no account, licence key, or registration.

How your documents are used. Parsing runs entirely on your own machine, in a local process. Document contents are used only to produce the outline, text slices, search results, and markdown you ask for, and are returned only to the caller. Documents are never uploaded to DocSlicer or to any third party. When running as an MCP server, DOCSLICER_MCP_ROOT bounds which directory tree may be read from and written to.

Where data is stored, and for how long. Parsed results are cached on your own disk — by default ~/.cache/docslicer-mcp, configurable with DOCSLICER_MCP_CACHE. The cache is pruned to a size ceiling (DOCSLICER_MCP_CACHE_MAX_MB, default 2048 MB); otherwise it persists until you delete it, and deleting the directory removes it permanently with no copy retained elsewhere. Nothing is written outside the cache directory and any output path you supply.

Network access and third parties. No network request is made for a local file. Requests leave your machine only when you pass an http(s) source: that URL is fetched directly from the host you named, and for HTML pages Playwright may load the subresources that page references, exactly as a browser would. Requests to sec.gov send a User-Agent header identifying the client, as the SEC fair-access policy requires. These hosts are third parties chosen by you, not by DocSlicer, and their own policies govern what they log. Set DOCSLICER_MCP_ALLOW_URLS=0 to reject remote sources entirely.

Third-party clients. When DocSlicer runs as an MCP server, the client (Claude, Cursor, …) handles the conversation under its own privacy policy. DocSlicer is not a party to that and receives nothing from it.

Contact. Privacy questions: jelle@docslicer.ai · Issues: https://github.com/DocSlicer/DocSlicer/issues

License

DocSlicer is dual-licensed:

• AGPL-3.0 — free to use, modify, and distribute, provided you comply with the AGPL's terms, including making the complete source of any application that uses DocSlicer available to its users (including over a network).
• Commercial license — for embedding DocSlicer in a closed-source or proprietary product, or offering it as part of a hosted/SaaS service without releasing your source.

See LICENSE-COMMERCIAL.md for details, or reach out about a commercial license.

mcp-name: io.github.DocSlicer/docslicer