docval 0.1.1

Validate and refactor Markdown documentation against source code using heuristics + LLM

docval

AI Cost Tracking

• 🤖 LLM usage: $0.1500 (1 commits)
• 👤 Human dev: ~$100 (1.0h @ $100/h, 30min dedup)

Generated on 2026-04-13 using openrouter/qwen/qwen3-coder-next

---

Validate and refactor Markdown documentation against source code — detect outdated, orphaned, duplicate, and invalid docs using heuristics + optional LLM.

How it works

docs/  ──→  chunk by heading  ──→  heuristic checks  ──→  cross-ref with code  ──→  (optional) LLM  ──→  report/fix

Three validation layers, each progressively deeper:

1. Heuristic validator (fast, free) — empty sections, broken internal links, TODO/FIXME markers, duplicate detection via difflib, stale version references, archive path detection, explicit deprecation markers
2. Cross-reference validator (fast, free) — checks that backtick-quoted symbols (ClassName, function_name), import paths in code blocks, and CLI commands actually exist in the project source
3. LLM validator (optional, paid) — semantic validation via litellm for chunks that heuristics couldn't resolve with high confidence

Installation

pip install docval

With LLM support:

pip install docval[llm]

From source:

git clone https://github.com/wronai/docval.git
cd docval
pip install -e ".[dev]"

CLI Usage

Scan and report issues

docval scan docs/
docval scan docs/ --project /path/to/repo -v
docval scan docs/ -o report.md
docval scan docs/ -o report.json

Fix documentation (dry-run by default)

docval fix docs/                           # preview changes
docval fix docs/ --no-dry-run              # apply fixes
docval fix docs/ --no-dry-run --llm        # with LLM validation

Generate a patch file

docval patch docs/ -o fixes.txt
docval patch docs/ --llm --model gpt-4o -o fixes.txt

View documentation statistics

docval stats docs/

LLM validation

export OPENAI_API_KEY=sk-...
docval scan docs/ --llm --model gpt-4o-mini
docval scan docs/ --llm --model anthropic/claude-sonnet-4-20250514
docval scan docs/ --llm --model groq/llama-3.3-70b-versatile

Any model supported by litellm works.

Python API

from pathlib import Path
from docval.pipeline import scan
from docval.reporters import ConsoleReporter, MarkdownReporter

# Run validation
result = scan(
    docs_dir=Path("docs/"),
    project_root=Path("."),
    use_llm=False,
)

# Print to console
ConsoleReporter(verbose=True).report(result)

# Write markdown report
MarkdownReporter().report(result, Path("validation-report.md"))

Using individual validators

from docval.chunker import chunk_directory
from docval.context import build_context
from docval.validators import HeuristicValidator, CrossRefValidator

# Chunk docs
doc_files = chunk_directory(Path("docs/"))

# Build project context
ctx = build_context(Path("."))

# Run heuristics
heuristic = HeuristicValidator(ctx=ctx)
heuristic.validate(doc_files)

# Cross-reference check
crossref = CrossRefValidator(ctx=ctx)
crossref.validate(doc_files)

# Inspect results
for f in doc_files:
    for chunk in f.chunks:
        if chunk.issues:
            print(f"{f.relative_path}:{chunk.line_start} [{chunk.status.value}] {chunk.heading}")
            for issue in chunk.issues:
                print(f"  {issue.severity.value}: {issue.message}")

What it detects

Check	Layer	Example
Empty sections	Heuristic	Heading with no body text
Broken internal links	Heuristic	[guide](./deleted-file.md)
Deprecated markers	Heuristic	DEPRECATED, OBSOLETE, DO NOT USE
Archive path	Heuristic	Files in docs/archive/ directories
Stale versions	Heuristic	References to v1.x when project is v3.x
Duplicates	Heuristic	>80% similar content across files
TODO/FIXME	Heuristic	Unfinished documentation markers
Orphaned code refs	CrossRef	`NonExistentClass` in backticks
Broken imports	CrossRef	from mypackage.deleted import X in code blocks
Semantic accuracy	LLM	Content that doesn't match actual project behavior

Architecture

src/docval/
├── cli.py                  # Click CLI: scan, fix, patch, stats
├── pipeline.py             # Orchestrates: discover → chunk → validate → report
├── models.py               # Data models: DocChunk, DocFile, ValidationResult
├── chunker.py              # MD → heading-based semantic chunks
├── context.py              # Build project context (AST, git, .toon files)
├── validators/
│   ├── heuristic.py        # Rule-based checks (free, fast)
│   ├── crossref.py         # Code ↔ docs cross-reference
│   └── llm_validator.py    # Semantic validation via litellm
├── actions/
│   └── executor.py         # Apply fixes: delete, archive, patch
└── reporters/
    ├── console.py           # Rich CLI output
    ├── markdown_report.py   # .md report
    └── json_report.py       # .json for CI/CD

Integration with .toon files

docval understands .toon.yaml files from the code2llm ecosystem. When present, it extracts module names, class names, and exported functions for cross-referencing, giving more accurate orphaned-reference detection.

License

Licensed under Apache-2.0.