SEC 10-K Structured Extraction Tool

Parses SEC EDGAR Form 10-K annual reports into standardized JSON. Automatically identifies the content and status of every Item — zero LLM cost, average latency < 1 second.

English | 中文

---

Features

• Full Item coverage: Parses all 16 Items across Part I–IV, outputting one of five statuses: extracted / incorporated_by_reference / not_applicable / reserved / missing
• Zero LLM cost: Purely rule-based pipeline — parsing itself takes ~0.03s with no API calls
• iXBRL support: Handles the iXBRL format mandatory for large filers since 2019
• XBRL financial extraction: Optionally reconstructs Item 8 primary financial statements directly from XBRL (income statement, balance sheet, cash flow statement, etc.)
• Sync / async API: Both Pipeline and AsyncPipeline are available

---

Quick Start

Installation

python -m venv .venv
source .venv/bin/activate  # Windows: .venv\Scripts\activate
pip install -r requirements.txt

Usage

Synchronous

from src.pipeline import Pipeline
from src.models import FilingInput

pipeline = Pipeline()

# Option 1: CIK + Accession Number
result = pipeline.run(FilingInput(
    cik="0000320193",
    accession_number="0000320193-23-000106",
))

# Option 2: Direct URL
result = pipeline.run(FilingInput(
    url="https://www.sec.gov/Archives/edgar/data/.../filing.htm",
))

# Save results (JSON + Markdown)
result = pipeline.run(input, save_to="output/")

Async

from src.async_pipeline import AsyncPipeline
from src.models import FilingInput
import asyncio

async def main():
    pipeline = AsyncPipeline()
    result = await pipeline.run_async(FilingInput(
        cik="0000320193",
        accession_number="0000320193-23-000106",
    ))

asyncio.run(main())

---

Output Format

{
  "filing_info": {
    "cik": "0000320193",
    "accession_number": "0000320193-23-000106",
    "company_name": "Apple Inc.",
    "fiscal_year_end": "2023-09-30"
  },
  "items": [
    {
      "part": "Part I",
      "item_number": "1",
      "item_title": "Business",
      "content_text": "Apple Inc. designs, manufactures...",
      "char_range": [1024, 45231],
      "status": "extracted"
    },
    {
      "part": "Part III",
      "item_number": "10",
      "item_title": "Directors, Executive Officers and Corporate Governance",
      "content_text": null,
      "char_range": null,
      "status": "incorporated_by_reference"
    }
  ]
}

---

How It Works

A purely rule-based pipeline — SEC regulations mandate Item numbering and ordering, giving the parser reliable anchors regardless of visual formatting variation:

Input (CIK + Accession Number or direct URL)
  ↓ fetch       Retrieve metadata and HTML from the SEC EDGAR API
  ↓ preprocess  HTML → plain text (handle iXBRL, tables, hyphenation)
  ↓ parse       RegexParser locates the start and end position of each Item
  ↓ postprocess Classify the status of each Item
Output (standardized JSON)

Architecture overview:

src/
├── pipeline.py              Main pipeline
├── async_pipeline.py        Async version
├── models.py                Data structures
├── patterns.py              Regex pattern definitions
├── postprocessor.py         Item status classification
├── item8_xbrl_facts.py      XBRL financial extraction
├── render_item8_markdown.py XBRL result renderer
├── parsers/
│   ├── regex_parser.py      Main parser
│   └── hybrid.py            Dispatcher (supports future LLM fallback)
└── eval/
    ├── metrics.py           Evaluation script
    └── runner.py            Batch evaluation entry point

---

Evaluation

Benchmarked on 35 filings, 12 companies, 2016–2026, with manually annotated ground truth:

Metric	Value
Status Accuracy	100.0% (788 / 788 items)
Critical Regressions	0
Content Length Normal Rate	99.0% (484 / 489 extracted items)
Head/Tail Match Pass Rate	Head 99.8% / Tail 100.0%
Average Latency	0.687s
LLM Cost	$0

In a separate large-scale test across 507 randomly sampled filings, structural error rate dropped 95.5% from the initial version.

Ground truth was manually annotated using SEC-10-K-Annotation-Tool. Full report: eval_datasets/results/.

Running Evaluation

python -m src.eval.metrics \
    --ground-truth eval_datasets/ground_truth \
    --output eval_datasets/results

---

XBRL Financial Extraction (Item 8)

Reconstructs Item 8 primary financial statements directly from XBRL, outputting a multi-period Markdown report.

from src.item8_xbrl_facts import get_item8_xbrl_facts
from src.render_item8_markdown import write_item8_markdown

payload = get_item8_xbrl_facts("0000019617", "0001628280-26-008131")
write_item8_markdown(payload, "output_item8.md")

Parses four XBRL source files (Instance Document, Presentation Linkbase, Label Linkbase, Schema) and classifies results into three blocks: main statements, numeric disclosures, and text disclosures.

---

Known Limitations

• Supports HTML-format filings from approximately 2000 onward. SGML / plain-text formats from before 1996 have no HTML structure and cannot be processed.

---

Contributing

PRs and Issues are welcome. If you encounter a filing that fails to parse, please open an Issue with the CIK and Accession Number.

---

License

MIT © LLMSystems