LangParse

Documents In, Knowledge Out.

LangParse is a vendor-neutral orchestration layer for document parsing and chunking in LLM / Agent applications — think LiteLLM, but for document parsing engines instead of LLM providers.

🚀 Project Status

LangParse is past the initial prototype: Markdown/DOCX/Excel/PDF parsing, semantic chunking, batch processing, quality checks, and a CI pipeline are all working end to end (649 tests passing). See docs/PROGRESS.md for the current module-by-module status and active roadmap — that file, not this section, is the source of truth for "what works today."

Still pre-1.0. Looking for early contributors and design partners, particularly to help wire up additional vertical engines (PaddleOCR-VL, vision-LLM backends) and pressure-test the engine-neutral routing design.

🤔 Why LangParse?

Document parsing tooling for RAG/Agent pipelines today falls into two camps, and neither solves the whole problem:

1. Single, opinionated parsing engines (MinerU, Docling, Marker, LlamaParse, ...). Each is strong within its own scope, but adopting one locks you into its trade-offs — switching between a lightweight generic parser and a heavyweight vertical engine (e.g. MinerU/DeepDoc for CJK or complex layouts) usually means rewriting your pipeline.
2. "Multi-engine" wrappers that aren't actually neutral. Projects like MegaParse or LiteParse nominally support several backends, but the product is structured to fund a flagship offering — MegaParse's own vision-based parser (its README benchmark table exists to show it beating the third-party engines it wraps), LiteParse's own local engine with an explicit upsell to LlamaParse for anything complex. Neither wires up self-hosted vertical engines like MinerU or DeepDoc as genuine peers.

LangParse is neither of those — it's the adapter/routing layer. One interface; generic engines (pdfplumber-based simple) and vertical/self-hosted engines (mineru and deepdoc today, paddle in progress) are equally first-class, pluggable backends, with no engine favored to drive adoption of a paid tier. Chunking strategy is a separate, independent choice on top of whichever engine parsed the document. Output is either the raw parsed document or chunked content — your call, same API.

Non-goals (kept here so scope doesn't drift):

• Not competing with MinerU / Docling / LlamaParse on raw extraction accuracy — that ceiling is set by the engine, not by this layer.
• Not a standalone parser-evaluation benchmark/leaderboard project (see OmniDocBench, SCORE-Bench for that). The fidelity scoring in services/fidelity.py exists to help you compare engines on your own documents when picking a backend — a supporting feature, not the product's identity.
• Not tied to any single vendor's cloud API as the only path — self-hosted engines and remote API engines are equally valid backends.

🏗️ Architecture

flowchart LR
    subgraph Input["Input Formats"]
        direction TB
        PDF["PDF"]
        DOCX["DOCX / DOC"]
        XLSX["XLSX / XLS / CSV"]
        MD["MD / TXT"]
    end

    subgraph Router["Router<br/>parsers/registry.py"]
        direction TB
        REG["content sniff first,<br/>extension as fallback<br/>(single source of truth)"]
    end

    subgraph GenericEngines["Generic Engines"]
        direction TB
        SIMPLE["simple<br/>(pdfplumber)"]
        DOCXP["DocxParser"]
        EXCELP["ExcelParser"]
        MDP["MarkdownParser"]
    end

    subgraph VerticalEngines["Vertical / Self-Hosted Engines"]
        direction TB
        MINERU["mineru ✅"]
        DEEPDOC["deepdoc ✅"]
        PADDLE["paddle 🚧 planned"]
        VISION["vision_llm 🚧 planned"]
    end

    subgraph Result["Unified Result"]
        direction TB
        PDR["ParsedDocumentResult<br/>pages / elements / tables / images"]
    end

    subgraph ChunkLayer["Chunking (pluggable)"]
        direction TB
        SEM["SemanticChunker<br/>blocks.py + semantic.py"]
    end

    subgraph Output["Output"]
        direction TB
        RAW["Raw parsed doc<br/>Markdown / JSON"]
        CHUNKS["Chunked content<br/>Chunk[] + metadata"]
    end

    subgraph Services["Services (cross-cutting)"]
        direction TB
        BATCH["batch_service"]
        QUALITY["quality checks"]
        BENCH["benchmark_service<br/>optional: compare engines<br/>on your own corpus"]
        METRICS["metrics"]
    end

    Input --> Router
    Router --> GenericEngines
    Router --> VerticalEngines
    GenericEngines --> Result
    VerticalEngines --> Result
    Result --> RAW
    Result --> SEM
    SEM --> CHUNKS
    Result -.-> Services

    style Input fill:#F5F5F5,color:#000000,stroke:#37D7FA,stroke-width:2px
    style Router fill:#F5F5F5,color:#000000,stroke:#8A8F98,stroke-width:2px
    style GenericEngines fill:#F5F5F5,color:#000000,stroke:#3E18F9,stroke-width:2px
    style VerticalEngines fill:#F5F5F5,color:#000000,stroke:#FF8705,stroke-width:2px
    style Result fill:#F5F5F5,color:#000000,stroke:#8A8F98,stroke-width:2px
    style ChunkLayer fill:#F5F5F5,color:#000000,stroke:#1FAA59,stroke-width:2px
    style Output fill:#F5F5F5,color:#000000,stroke:#FF8DF2,stroke-width:2px
    style Services fill:#FAFAFA,color:#000000,stroke:#8A8F98,stroke-width:1px,stroke-dasharray: 4 3

Same shape as the LiteParse diagram, different point: theirs shows a single engine's internal pipeline (format conversion → text extraction → OCR → grid projection). This one shows the routing layer around multiple engines — generic and vertical engines are peers feeding the same ParsedDocumentResult, chunking is a separate optional stage on top, and services (batch/quality/benchmark) cut across the whole pipeline rather than living inside any one engine.

✨ Core Features

• 🔌 Engine-neutral routing: Generic (simple) and vertical (mineru, deepdoc, with paddle in progress) PDF engines share one interface and one output shape (ParsedDocumentResult). No default "flagship" engine — you pick based on your documents.
• 📄 Multi-format parsing: .pdf .docx .doc .xlsx .xlsm .xls .csv .md .txt out of the box, all normalized to the same structured result.
• 📗 Lossless OOXML facts: .xlsx/.xlsm parsing preserves coordinates, raw/display values, formulas and cached values, merges, style fingerprints, visibility, dimensions, print areas, comments, hyperlinks, and object anchors in WorkbookIR.snapshot.
• 🧩 Pluggable semantic chunking: Markdown-structure-aware chunking (headings, lists, tables, code blocks), decoupled from which engine produced the content.
• 📡 Unified output: Get the parsed document as-is, or chunked with rich metadata (source_file, page_number, header, ...) — same API either way.
• 📊 Optional fidelity scoring: services/fidelity.py plus the benchmark CLI command let you quantitatively compare engines on your own documents when you need evidence for a choice.

📦 Installation

The first public release candidate is 0.1.0rc1:

pip install --pre "langparse==0.1.0rc1"

Install only the optional capabilities you need:

pip install "langparse[excel]"
pip install "langparse[excel,model]"  # optional OpenAI workbook disambiguation
pip install "langparse[deepdoc]"
pip install "langparse[all]"

Calling an existing remote MinerU API needs only the core package. Install langparse[mineru] only when this Python environment must provide and start a local mineru-api orchestrator. Local inference backends remain an explicit operator choice; for example, install the official mineru[pipeline] extra for the CPU/GPU pipeline backend instead of pulling every platform backend.

⚡ Quick Start (Alpha)

You can try the current alpha version by cloning the repository:

git clone https://github.com/syw2014/langparse.git
cd langparse
pip install -e .

Basic Usage

from langparse import MarkdownParser, SemanticChunker

# 1. Initialize
parser = MarkdownParser()
chunker = SemanticChunker()

# 2. Parse a file (currently supports .md)
doc = parser.parse("README.md")

# 3. Chunk it semantically
chunks = chunker.chunk(doc)

# 4. Inspect chunks
for chunk in chunks:
    print(f"Header Path: {chunk.metadata.get('header_path')}")
    print(f"Content: {chunk.content[:50]}...")

Chunking

Chunks respect a size budget while following Markdown structure. Sections come from headings; within a section, blocks are packed up to max_chunk_size.

SemanticChunker(max_chunk_size=1000, overlap=0, length_function=len)

• length_function measures chunk size. The default counts characters and pulls in no dependencies; pass a tokenizer's encoder to budget in tokens:

  import tiktoken
  encoder = tiktoken.get_encoding("cl100k_base")
  SemanticChunker(max_chunk_size=512, length_function=lambda t: len(encoder.encode(t)))
  

• overlap is off by default. It duplicates content into the vector store, so it is opt-in.
• Tables that exceed the budget split by row with the header row repeated in each part, so every chunk stays readable on its own.
• Code blocks are never split — splitting would leave unterminated fences. An oversized one emits whole with oversized: True in its metadata.
• A # inside a fenced code block is not treated as a heading.

Each chunk carries header, header_level, header_path, page_numbers and chunk_index.

From the CLI, --chunk adds a chunks array to JSON output (and separates chunks with --- in Markdown output), and activates the chunk metrics:

langparse parse paper.pdf --chunk --format json
langparse parse docs/ --batch --chunk --metrics --output-dir out

Structured Excel results

OOXML workbooks are not treated as paginated pandas tables. Each sheet keeps a stable compatibility ordinal, while the result sets paginated=False and exposes lossless source facts, deterministic logical tables, coverage diagnostics, semantic Markdown, and source-aware table-row chunks from one parse:

from langparse.services.parse_service import ParseService

parsed = ParseService().parse_result(
    "budget.xlsx",
    chunk=True,
    chunk_profile="retrieval",
)
analysis_chunks = ParseService().chunk_result(
    parsed,
    chunk_profile="analysis",
)
print(parsed.structure.snapshot.sheets[0].cells["B2"].formula)
print(parsed.diagnostics.coverage_ratio)
print([block.kind for block in parsed.structure.sheets[0].blocks])
print(parsed.diagnostics.source_ref_validity_ratio)
print(parsed.chunks[0].metadata["chunk_type"])
print(parsed.chunks[0].metadata["source_ranges"])
print(analysis_chunks[0].structured_payload.get("records"))

sheet_table = parsed.structure.sheets[0].blocks[0].logical_table
cross_sheet_table = parsed.structure.table_continuations[0].logical_table

The parser deterministically separates tables across blank row/column bands, merges repeated print fragments, builds multi-level header paths, classifies sections/data/totals, and chunks complete logical rows without crossing section boundaries. Candidate regions are conservatively classified as logical tables, forms, matrices, text, or explicit unclassified raw grids; every kind has a source-aware Markdown and chunk path. structure.snapshot and compatibility tables retain the original cell-level view. High-confidence adjacent-Sheet continuations expose one aggregate logical table through structure.table_continuations; insufficient evidence keeps tables independent and records an ambiguous or rejected diagnostic. Markdown and chunks remain source-Sheet based rather than duplicating the aggregate, and source-member chunks can be regrouped by continuation_id. Retrieval/analysis dual chunk profiles are built into the library, batch service, and CLI. retrieval is the default profile and uses a 1000-character budget; analysis uses 4000. Both preserve complete rows and exact source references. Analysis chunks add normalized, source-linked records while keeping cell-level facts, including formulas and cached values, in structure.snapshot. A parsed result can generate another profile repeatedly with chunk_result() without reparsing or mutating its structure. The analysis profile is only available for OOXML workbook results: CSV, legacy .xls, and non-workbook inputs keep their compatibility paths. Use structure.snapshot for exact cell or formula analysis rather than treating analysis chunks as a replacement for the fact layer.

langparse parse budget.xlsx --chunk --chunk-profile analysis --format json

Optional workbook model disambiguation

Workbook model disambiguation remains explicitly opt-in. The default is off: constructing ExcelParser() or calling ParseService without workbook_disambiguation performs no model Adapter or cache construction, reads no provider configuration, and creates no implicit model network work. Install the official OpenAI SDK integration separately from the core parser:

pip install "langparse[excel,model]"
export OPENAI_API_KEY="..."
export OPENAI_MODEL="gpt-4o-mini"
# Optional for an OpenAI-compatible endpoint:
export OPENAI_BASE_URL="https://example.invalid/v1"

langparse parse budget.xlsx --model --disambiguation auto --format json

API keys are intentionally not accepted as CLI arguments because process arguments and shell history are not secret stores. --model or an explicit --disambiguation auto|required enables network work; environment variables alone never enable it. Set LANGPARSE_DISABLE_MODEL=1 for the runtime kill switch.

The library interface may either use the built-in OpenAIWorkbookStructureAdapter or inject another WorkbookStructureModelAdapter:

from langparse.parsers.excel_parser import ExcelParser
from langparse.services.parse_service import ParseService
from langparse.workbooks.modeling import WorkbookDisambiguation

# `adapter` is supplied by the caller and implements
# WorkbookStructureModelAdapter.
direct = ExcelParser(
    disambiguation=WorkbookDisambiguation.auto(adapter)
).parse_result("budget.xlsx")

strict = ParseService().parse_result(
    "budget.xlsx",
    workbook_disambiguation=WorkbookDisambiguation.required(adapter),
)

Phase 4A is limited to choice-only region-kind disambiguation. Only a locally ambiguous, unclassified region with at least two compatible registered kinds is eligible. A response can only be selected with that case's registered case_id + choice_id, or abstained; it cannot express a value, formula, coordinate, range, header, row role, continuation, or arbitrary structure. Provider-reported confidence is diagnostic only. The selected kind is applied from the retained workbook snapshot and must still pass local materialization, coverage, reconstruction, row-conservation, continuation, and source-reference validation.

Model application is workbook-atomic. If any attempted selection cannot be materialized, or the tentative workbook fails a continuation or structural validator, every attempted selection is restored to its retained deterministic block and all validators run again. required reports every reverted case as unresolved.

auto keeps the deterministic local fallback and records sanitized diagnostics when the provider, cache, limits, response contract, materialization, or final validation fails, or when the provider abstains. required raises RequiredWorkbookDisambiguationError for unresolved eligible ambiguity and the typed error passes through ExcelParser and ParseService; a workbook with no eligible ambiguity succeeds with zero calls in either mode.

An enabled WorkbookDisambiguation value owns a private, thread-safe, process-local runtime/cache. Reusing that same value across ExcelParser, ParseService, or batch calls allows a validated response to become a re-decoded cache hit; off constructs no runtime or cache. max_cases limits cases considered, while max_calls is a workbook-wide hard budget of actual Adapter invocations, including retries; cache hits consume no calls. Policy timeouts must be finite positive non-boolean real values, and count/byte limits must be exact positive non-boolean integers.

The candidate request is deliberately narrow. It can include the target Sheet name and source range, visible cell coordinates and display text, value type, style fingerprint, merge geometry, local scalar features, and the registered choices for that region. It omits hidden Sheets, formulas and cached formula values, comments, hyperlinks, images, other regions, credentials, and provider secrets. If any cell in the complete candidate envelope contains a formula—even an unlisted cell or merged child—the whole case is locally unavailable and no formula or cached result is projected. Cell text is treated as untrusted Prompt Injection data: the Adapter port exposes no tool channel, and exact response fields, request checksum, case/choice membership, size limits, and local validation prevent cell instructions from expanding the operation. Duplicate JSON member names are rejected at every response-object depth. Diagnostics do not retain prompts, cell text, raw replies, or provider exception messages. The process-local, non-persistent cache has a narrower but different contract: it retains only response envelope bytes that have already passed strict response decoding, and every hit is decoded and validated again. Nothing is written to disk, but provider-supplied strings inside that envelope may remain in process memory until the owning disambiguation value and its private runtime are released. Each model-call audit records local schema, prompt, rule, validator, and privacy versions plus the deterministic fallback rule confidence; these values never come from the provider.

The workbook ambiguity evaluator is available with:

langparse eval \
  samples/workbook_ambiguity/public-manifest.json \
  --output-dir reports/workbook-ambiguity

# Live provider evidence is still explicit:
langparse eval private-manifest.json --model

Reports are immutable by digest and reject incomplete or modified replays. production_ready additionally requires holdout data, at least 30 ambiguous cases, and separate operational staging evidence; the bundled tuning seed can never satisfy that release gate by itself.

Cost circuit breakers never infer prices from a model name. Library callers that set max_cost_usd_per_workbook must also supply input_cost_usd_per_million, output_cost_usd_per_million, and a stable cost_pricing_version in WorkbookModelPolicy. These rates should come from the deployment's provider contract, including for OpenAI-compatible endpoints.

Phase 4B includes the optional OpenAI SDK Adapter, environment-based provider configuration, a strict structured-response contract, observed usage/cost circuit breakers, and an immutable evaluation report pipeline. This is a usable provider path, but not production-effectiveness evidence by itself: release still requires a representative private holdout, staging latency/cost and failure-mode evidence, and a provider privacy review. The observed token and cost budgets stop retries or later calls after reported usage reaches the limit; they cannot prevent the first provider call from exceeding a budget and therefore are circuit breakers rather than billing guarantees.

Summary/index chunks, rich .xls/.xlsb adapters, image/chart semantic blocks, standard bundle output, and further production hardening remain follow-up work; screenshots and VLM are Phase 4C, a second domain contract is Phase 4D, and delimited and legacy inputs keep the compatibility adapter for now.

Scanned PDFs

The simple engine falls back to OCR when a page turns out to be an image. Detection needs both a page-covering image and a thin text layer — a scanned page often carries a watermark, and that watermark text is enough to clear any threshold low enough to avoid firing on genuinely sparse text pages.

pip install "langparse[ocr]"

PDFParser(engine="simple", enable_ocr=True, ocr_min_chars=500)

Pages that took the fallback report ocr_applied and ocr_text_chars in their metadata, which surface in ParseMetrics. Without rapidocr_onnxruntime installed the parse still succeeds — the page simply keeps whatever text layer it had, rather than failing.

Measuring parse fidelity

Quality checks measure structure: page counts, whether any tables were found. They say nothing about whether the content is correct. To measure that, give a benchmark sample a reference:

{
  "id": "report-01",
  "path": "samples/report.pdf",
  "expected_markdown": "samples/report.expected.md",
  "expected_tables": [[["Header A", "Header B"], ["1", "2"]]]
}

• Text is scored by word-level normalised edit distance. Words rather than characters, because a reflowed line break is not an error but a dropped word is.
• Tables are scored by TEDS. Cell substitution costs the normalised character distance between the two cells, so a typo scores better than a wrong value, and a dropped row costs more than a changed cell.

Samples without a reference are reported as unscored, never as perfect.

MinerU Runtime

LangParse can run MinerU through mineru-api.

Runtime selection works like this:

• If you pass or configure api_url, LangParse calls that MinerU service directly.
• A remote mineru-api backed by a separate vLLM server also receives backend and server_url as /file_parse form fields. This path does not require the local [mineru] extra.
• If api_url is not set, LangParse will try to start a local mineru-api service and manage its lifecycle for the current parse.
• If mineru-api is not installed, pass --auto-install-runtime or auto_install_runtime=True to let LangParse install the configured runtime package in the current Python environment before starting the local service.

You can still control CPU/GPU selection and model/download directories through runtime parameters or configuration.

For local managed services:

• model_dir means "use this already-downloaded MinerU model directory"
• download_dir becomes the MinerU home root used by the local service, so MinerU will keep its default cache/config layout under that directory
• model_policy="require_existing" disables first-run download fallback and requires an existing local model setup

from langparse import AutoParser

doc = AutoParser.parse(
    "paper.pdf",
    engine="mineru",
    api_url="http://mineru.example:25820",
    backend="vlm-http-client",
    server_url="http://vlm.example:21670",
    request_timeout=900,
)

from langparse import AutoParser

cpu_doc = AutoParser.parse(
    "paper.pdf",
    engine="mineru",
    device="cpu",
    download_dir="./downloads",
)

from langparse import AutoParser

local_doc = AutoParser.parse(
    "paper.pdf",
    engine="mineru",
    model_dir="./preloaded-models",
    model_policy="require_existing",
)

Environment variables:

export LANGPARSE_MINERU_API_URL=http://127.0.0.1:8000
export LANGPARSE_MINERU_BACKEND=vlm-http-client
export LANGPARSE_MINERU_SERVER_URL=http://vlm.example:21670
export LANGPARSE_MINERU_REQUEST_TIMEOUT=900
export LANGPARSE_MINERU_DEVICE=cuda
export LANGPARSE_MINERU_MODEL_DIR=./models
export LANGPARSE_MINERU_DOWNLOAD_DIR=./downloads
export LANGPARSE_MINERU_MODEL_POLICY=require_existing
export LANGPARSE_MINERU_AUTO_INSTALL_RUNTIME=true

CLI Examples

The CLI handles every supported format, not just PDF. --engine applies to PDFs only; other formats route to their own parser automatically:

langparse parse report.docx --format json
langparse parse notes.md --output notes.out.md
langparse parse mixed_folder/ --batch --output-dir out --metrics

Supported extensions: .pdf, .docx, .doc, .xlsx, .xlsm, .xls, .csv, .md, .txt. Batch directory expansion picks up all of them; unsupported files exit with code 2 and a one-line message.

Single-file parsing:

langparse parse paper.pdf --engine mineru \
  --api-url http://mineru.example:25820 \
  --mineru-backend vlm-http-client \
  --mineru-server-url http://vlm.example:21670 \
  --mineru-request-timeout 900 \
  --format json

Batch parsing:

langparse parse docs/ --engine mineru --batch --output-dir out --format json

Batch parsing with lightweight metrics and skip-existing behavior:

langparse parse docs/ --engine mineru --batch --output-dir out --format json --max-workers 4 --skip-existing --metrics

Run a product-readiness benchmark:

langparse benchmark samples/public.example.json --engine mineru --output-dir reports --max-workers 2

Benchmark reports include success rate, elapsed time, pages per second, table counts, OCR indicators, reading-order warnings, header/footer filtering counts, and image/caption metadata coverage.

If you want LangParse to manage a local MinerU service, omit --api-url. You can also override the local launch command and bind address:

langparse parse paper.pdf --engine mineru --api-command "mineru-api" --api-host 127.0.0.1 --api-port 8000

Install MinerU automatically in the current Python environment if mineru-api is missing:

langparse parse paper.pdf --engine mineru --auto-install-runtime --device cpu --format json

Use an existing local model directory without allowing implicit downloads:

langparse parse paper.pdf --engine mineru --model-dir ./preloaded-models --model-policy require_existing

🛠️ Development & Local Testing

LangParse uses uv for environment and dependency management. The checked-in .venv is uv-managed and intentionally has no pip, so run everything through uv run (a bare pip/python on your shell may resolve to a different interpreter, e.g. Anaconda).

Set up the environment

# Install all dependencies (including dev/test) from uv.lock
uv sync --all-extras

# Or install just what you need
uv sync                      # core only (no third-party dependencies)
uv pip install -e ".[pdf]"   # PDF parsing (pdfplumber)
uv pip install -e ".[docx]"  # Word parsing (python-docx)
uv pip install -e ".[excel]" # Excel parsing (pandas + openpyxl)
uv pip install -e ".[model]" # Optional OpenAI workbook disambiguation
uv pip install -e ".[ocr]"   # OCR (rapidocr_onnxruntime)
uv pip install -e ".[mineru]"# MinerU API/orchestrator (local backend is explicit)
uv pip install -e ".[deepdoc]"# DeepDoc runtime (OCR/layout/table ONNX weights, ~100MB download on first run)
uv pip install -e ".[all]"   # everything above

Note: the core install has no third-party dependencies. The PDF/DOCX/Excel parsers require the optional extras above; without them a parse fails with an ImportError naming the missing package rather than crashing. pip install -e ".[dev]" is enough to run the test suite.

Run the tests

uv run pytest -q

Smoke-test locally

# Markdown parse + semantic chunk (no extra deps needed)
uv run python examples/basic_usage.py

# Parse a PDF of your own (requires the [pdf] extra)
uv run langparse parse your.pdf --engine simple --format json

# Run the benchmark on the bundled manifest template
uv run langparse benchmark samples/public.example.json --engine simple --output-dir reports

The repository ships samples/public.example.json as a benchmark manifest template. data/ is where local test documents go; it is git-ignored, so bring your own.

💬 Contact

For questions, feature requests, or bug reports, the preferred method is to open an issue on this GitHub repository. This allows for transparent discussion and helps other users who might have the same question.

Citing LangParse

If you use LangParse in your research, product, or publication, we would appreciate a citation! You can use the following BibTeX entry:

@software{LangParse_2025,
  author = {syw2014},
  title = {LangParse: A universal document parsing and text chunking engine for LLM or agent applications},
  month = {November},
  year = {2025},
  publisher = {GitHub},
  url = {https://github.com/syw2014/langparse}
}

Changelog

See CHANGELOG.md (中文) for release notes and the dated development history.

License

This project is licensed under the Apache 2.0 License.