LLM Documentation Scraper (doc-scraper)

A configurable, concurrent, and resumable web crawler written in Go. Specifically designed to scrape technical documentation websites, extract core content, convert it cleanly to Markdown format suitable for ingestion by Large Language Models (LLMs), and save the results locally.

Overview

This project provides a powerful command-line tool to crawl documentation sites based on settings defined in a config.yaml file. It navigates the site structure, extracts content from specified HTML sections using CSS selectors, and converts it into clean Markdown files.

Why Use This Tool?

• Built for LLM Training & RAG Systems - Creates clean, consistent Markdown optimized for ingestion
• Preserves Documentation Structure - Maintains the original site hierarchy for context preservation
• Production-Ready Features - Offers resumable crawls, rate limiting, and graceful error handling
• High Performance - Uses Go's concurrency model for efficient parallel processing

Goal: Preparing Documentation for LLMs

The main objective of this tool is to automate the often tedious process of gathering and cleaning web-based documentation for use with Large Language Models. By converting structured web content into clean Markdown, it aims to provide a dataset that is:

• Text-Focused: Prioritizes the textual content extracted via CSS selectors
• Structured: Maintains the directory hierarchy of the original documentation site, preserving context
• Cleaned: Converts HTML to Markdown, removing web-specific markup and clutter
• Locally Accessible: Provides the content as local files for easier processing and pipeline integration

Key Features

Getting Started

Prerequisites

• Go: Version 1.26 or later
• Git: For cloning the repository
• Disk Space: Sufficient for storing crawled content and state database

Installation

Option 1: Direct Installation (Recommended)

Install the latest version directly from GitHub:

go install github.com/Sriram-PR/doc-scraper/v2/cmd/doc-scraper@latest

This installs the doc-scraper binary to your GOPATH/bin directory (usually ~/go/bin or %USERPROFILE%\go\bin). Make sure this directory is in your PATH.

Option 2: Clone and Build

1. Clone the repository:

   bash

   Copy

   git clone https://github.com/Sriram-PR/doc-scraper.git
   cd doc-scraper
   

2. Install Dependencies:

   bash

   Copy

   go mod tidy
   

3. Build the Binary:

   bash

   Copy

   make build
   # or: go build -o doc-scraper ./cmd/doc-scraper
   

   This creates an executable named doc-scraper in the project root.

Quick Start

Create a minimal config.yaml in the project root:

output_base_dir: "./crawled_docs"
state_dir: "./crawler_state"
enable_jsonl_output: true
sites:
  rust_cli_book:
    start_urls:
      - "https://rust-cli.github.io/book/index.html"
    allowed_domain: "rust-cli.github.io"
    allowed_path_prefix: "/book/"
    content_selector: "#content, main"
    max_depth: 2          # seed plus one level; set 0 for the whole book

Run the crawl:

./doc-scraper crawl -site rust_cli_book -loglevel info

The Markdown, plus pages.jsonl, llms.txt, and llms-full.txt, lands under ./crawled_docs/rust_cli_book/ (output is organized by site key). A small book like this finishes in a few seconds; large sites can take minutes, so start with a low max_depth to gauge size before removing the bound.

Configuration (config.yaml)

A config.yaml file is required to run the crawler. Create this file in the project root or specify its path using the -config flag.

Key Settings for LLM Use

When configuring for LLM documentation processing, pay special attention to these settings:

• sites.<your_site_key>.content_selector: Define precisely to capture only relevant text
• sites.<your_site_key>.allowed_domain / allowed_path_prefix: Define scope accurately
• skip_images: Images are not downloaded by default (text-first). Set to false globally or per-site to download and localize images for offline consumption
• Adjust concurrency/delay settings based on the target site and your resources

Example Configuration

# Global settings (applied if not overridden by site)
default_delay_per_host: 500ms
num_workers: 8
num_image_workers: 8
max_requests: 48
max_requests_per_host: 4
output_base_dir: "./crawled_docs"
state_dir: "./crawler_state"
max_retries: 4
initial_retry_delay: 1s
max_retry_delay: 30s
global_crawl_timeout: 0s
skip_images: true # Default. Set to false to download and localize images
max_image_size_bytes: 10485760 # 10 MiB (applies only when images are downloaded)
enable_jsonl_output: true
jsonl_output_filename: "pages.jsonl"

# HTTP Client Settings
http_client_settings:
  timeout: 45s
  max_idle_conns_per_host: 6

# Site-specific configurations
sites:
  # Key used with -site flag
  pytorch_docs:
    start_urls:
      - "https://pytorch.org/docs/stable/"
    allowed_domain: "pytorch.org"
    allowed_path_prefix: "/docs/stable/"
    content_selector: "article.pytorch-article .body"
    max_depth: 0 # 0 for unlimited depth
    skip_images: false # Opt in to downloading images for this site
    disallowed_path_patterns:
      - "/docs/stable/.*/_modules/.*"
      - "/docs/stable/.*\.html#.*"

  tensorflow_docs:
    start_urls:
      - "https://www.tensorflow.org/guide"
      - "https://www.tensorflow.org/tutorials"
    allowed_domain: "www.tensorflow.org"
    allowed_path_prefix: "/"
    content_selector: ".devsite-article-body"
    max_depth: 0
    delay_per_host: 1s  # Site-specific override
    # Disable JSONL output for this site, overriding global
    enable_jsonl_output: false
    disallowed_path_patterns:
      - "/install/.*"
      - "/js/.*"

Full Configuration Options

HTTP Client Settings: (Global; cannot be overridden per site. Pool, dialer, and TLS timings are baked into pkg/fetch with sane defaults and are not exposed as config knobs.)

• timeout: Overall request timeout (default 45s)
• max_idle_conns_per_host: Idle connections per host (default 2)
• allow_private_networks: Disables the SSRF guard that blocks dials to loopback / private / link-local / CGNAT / multicast addresses. Default false. Set to true only if you intentionally crawl internal documentation servers reachable via private IPs.

Site-Specific Configuration Options:

• start_urls: Array of starting URLs for crawling (Required)
• allowed_domain: Restrict crawling to this domain (Required)
• allowed_path_prefix: Restrict crawling to URLs under this path prefix (Optional; defaults to /, the whole domain). Setting it is strongly recommended to bound scope
• content_selector: CSS selector for main content extraction, or "auto" for automatic detection (Required)
• max_depth: Exclusive upper bound on crawl depth from start URLs. Start pages are depth 0, so 1 crawls only the start pages, 2 adds their directly-linked pages, and so on. 0 = unlimited. URLs discovered from a sitemap.xml are seeded at depth 1 (one hop from the site root), so they are still bounded by max_depth: max_depth: 1 stays start-only and skips sitemap expansion
• delay_per_host: Override global delay setting for this site
• disallowed_path_patterns: Array of regex patterns for URLs to skip
• link_extraction_selectors: Array of CSS selectors for additional link extraction areas
• respect_nofollow: Boolean. Whether to respect rel="nofollow" links
• user_agent: String. Override global user agent for this site
• skip_images: Override the global image setting for this site. Images are skipped unless this (or the global skip_images) is set to false
• max_image_size_bytes: Integer. Override global max image size for this site
• allowed_image_domains: Array of domains from which to download images
• disallowed_image_domains: Array of domains to block image downloads from
• enable_jsonl_output: true or false. Override global JSONL output enablement for this site
• jsonl_output_filename: String. Override global JSONL output filename for this site

Usage

Execute the compiled binary from the project root directory:

./doc-scraper <command> [options]

Commands

Command Options

crawl:

Note: One of -site, -sites, or --all-sites is required.

add:

doc-scraper add https://vitepress.dev/guide/what-is-vitepress

Probes the site with a handful of polite requests (the page, robots.txt, llms.txt, the sitemap), then shows what it found before anything is written: the detected framework and content selector (validated against the fetched page), a crawl scope clustered from the sitemap with page counts as evidence, sibling version/locale trees proposed as exclusions, and a markdown preview of the extracted page with code-block fidelity numbers. The entry is appended to your config only after you confirm; the rest of the file is preserved byte-for-byte, comments included.

Exit codes: 0 written, 1 error, 2 drafted but not written. For agents and scripts: add -dry-run -json <url> inspects, then add -yes <url> commits; with no terminal attached the command fails fast instead of waiting on stdin. Sites whose robots.txt disallows crawling the given path are refused, and robots rules that restrict AI crawlers are surfaced as a warning.

config validate:

config list:

mcp-server: (stdio transport only; the SSE transport was removed in v2.x)

watch:

Note: One of -site, -sites, or --all-sites is required.

Example Usage Scenarios

Basic Crawl:

./doc-scraper crawl -site tensorflow_docs -loglevel info

Resume a Large Crawl:

./doc-scraper crawl -site pytorch_docs --resume -loglevel info

Validate Configuration:

./doc-scraper config validate -config config.yaml
./doc-scraper config validate -site pytorch_docs  # Validate specific site

List Available Sites:

./doc-scraper config list

High Performance Crawl with Profiling:

./doc-scraper crawl -site small_docs -loglevel warn -pprof localhost:6060

Debug Mode for Troubleshooting:

./doc-scraper crawl -site test_site -loglevel debug

Parallel Crawl of Multiple Sites:

./doc-scraper crawl -sites pytorch_docs,tensorflow_docs,langchain_docs

Crawl All Configured Sites:

./doc-scraper crawl --all-sites

Start MCP Server for Claude Desktop:

./doc-scraper mcp-server -config config.yaml

Incremental Crawling

crawl -incremental (which implies --resume, and is also what watch mode uses) re-fetches every previously-crawled page and re-checks it for changes:

• Change detection is content-scoped: it hashes the extracted content-selector region, not the raw page. Churn in the page shell (navigation, analytics, build timestamps, CSRF tokens) outside the content selector does not count as a change.
• Pages whose content region is unchanged are skipped without re-converting, re-downloading images, or rewriting output.
• Pages whose content region changed are fully reprocessed and their output is rewritten.
• A page that now returns an error (e.g. 404) on re-crawl leaves its previously-crawled output as-is; nothing is pruned.

Because there is no conditional-request support yet, incremental mode still performs the HTTP fetch for each known page; the savings come from skipping the downstream processing of unchanged pages.

Output Structure

Crawled content is saved under the output_base_dir defined in the config, organized by site key and preserving the site structure. Keying by site key (rather than domain) keeps two site configs that target the same domain in separate trees:

<output_base_dir>/
└── <sanitized_site_key>/            # e.g., flask_docs
    ├── images/                       # Always created; only populated when skip_images: false
    │   ├── image1.png
    │   └── image2.jpg
    ├── index.md                      # Markdown for the root path
    ├── <jsonl_output_filename>       # If enable_jsonl_output: true
    ├── llms.txt                      # Manifest of pages (auto-generated, when JSONL is enabled)
    ├── llms-full.txt                 # Full content concatenated (auto-generated, when JSONL is enabled)
    ├── topic_one/
    │   ├── index.md
    │   └── subtopic_a.md
    └── topic_two.md

llms.txt and llms-full.txt

When JSONL output is enabled, the crawler also emits llms.txt and llms-full.txt following the llmstxt.org convention. llms.txt is a markdown manifest (H1 + summary blockquote + ## Pages list of every crawled page with title and URL). llms-full.txt concatenates the full markdown content of every page, with section separators. Both files are regenerated on every crawl from the JSONL source of truth, so resumed crawls produce a complete updated manifest.

Output Format

Each generated Markdown file begins with a YAML frontmatter block carrying page metadata, followed by the converted content:

• YAML frontmatter (delimited by ---) with title, url (source URL), crawled_at (RFC3339 timestamp), content_hash (SHA-256 of the content, matching the JSONL record), and depth
• Clean content converted from HTML to GitHub-Flavored Markdown, preserving tables
• Relative links to other pages (when within the allowed domain)
• Local image references (if images are enabled)

Example:

---
title: 'Authentication'
url: https://docs.example.com/api/auth
crawled_at: "2026-08-09T12:00:00Z"
content_hash: 9f2b...c1a4
depth: 2
---

# Authentication

...page content as Markdown...

JSONL Output

When enabled, the crawler writes one JSON object per line to a JSONL file. This format is designed for ingestion into RAG pipelines and downstream indexers.

Enable it:

enable_jsonl_output: true
jsonl_output_filename: "pages.jsonl"  # default

The file mixes two record kinds, distinguished by the record_type field:

• page records, one per crawled page.
• A single crawl_meta record as the final line, holding the crawl-level summary. Resuming rewrites the file to drop any leftover crawl_meta record before appending a fresh one at close, so a closed file always contains exactly one crawl_meta record.

page record fields (from PageJSONL):

crawl_meta record fields (from CrawlMetaJSONL):

The output file is written to each site's output directory. Both the enable flag and filename can be overridden per site.

Auto Content Detection

When you set content_selector: "auto" for a site, the crawler automatically detects the documentation framework and applies the appropriate content selector.

Supported Frameworks

Detection recognizes 30+ documentation generators and hosted platforms, checked in three tiers of decreasing trust: the <meta name="generator"> tag, structural DOM signatures (attributes, ids, classes), and asset path patterns. Covered families include Docusaurus, VitePress, VuePress, Starlight/Astro, Nextra, Fumadocs, Mintlify, GitBook, MkDocs (Material, ReadTheDocs theme, and plain), Sphinx (furo, pydata, book, RTD, and classic themes), Antora, Docsy, hugo-book, Geekdoc, just-the-docs, mdBook, rustdoc, pkg.go.dev, Javadoc, Doxygen, TypeDoc, Writerside, ReadMe.com, Intercom, and Docus.

Every detected selector is validated against the live page before it is trusted: if it matches nothing or captures too little text, the crawler falls back instead of extracting empty content. Client-rendered shells (Docsify, Swagger UI, Redoc, Scalar, Document360, and generic empty-body SPAs) are recognized and reported as needing JavaScript rendering rather than silently producing an empty crawl.

Fallback Behavior

If no known framework is detected (or the detected selectors do not match the page), the crawler uses Mozilla's Readability algorithm to extract the main content. This works well on classic server-rendered docs, but can drop code blocks on some modern sites, so doc-scraper add's preview reports code-block fidelity before you commit a config.

Example Usage

sites:
  pytorch_docs:
    start_urls:
      - "https://pytorch.org/docs/stable/"
    allowed_domain: "pytorch.org"
    allowed_path_prefix: "/docs/stable/"
    content_selector: "auto"  # Auto-detect framework
    max_depth: 0

Parallel Site Crawling

Crawl multiple documentation sites concurrently with shared resource management. The orchestrator coordinates multiple crawlers while respecting global rate limits and semaphores.

Usage

# Crawl specific sites in parallel
./doc-scraper crawl -sites pytorch_docs,tensorflow_docs,langchain_docs

# Crawl all configured sites
./doc-scraper crawl --all-sites

# Resume parallel crawl
./doc-scraper crawl -sites pytorch_docs,tensorflow_docs --resume

Resource Sharing

When running parallel crawls, the following resources are shared across all site crawlers:

• Global semaphore: Limits total concurrent requests across all sites
• HTTP client: Shared connection pooling
• Rate limiter: Respects per-host delays

Each site still maintains its own:

• BadgerDB store for state persistence
• Output directory for crawled content
• Per-host semaphores for domain-specific limiting

Results Summary

After all sites complete, the orchestrator outputs a summary:

===========================================
Parallel crawl completed in 2m30s
Site Results:
  pytorch_docs: SUCCESS - 1500 pages in 1m20s
  tensorflow_docs: SUCCESS - 2000 pages in 2m15s
  langchain_docs: FAILED - 0 pages in 3s
    Error: initial fetch failed for start URL (see logs)
-------------------------------------------
Total: 3 sites (2 success, 1 failed), 3500 pages processed
===========================================

Unknown or misspelled site keys are rejected before the crawl starts, so they never appear as a FAILED row in this summary. For example, crawl -sites pytorch_docs,typo_key exits immediately (non-zero) with:

Invalid site keys: site 'typo_key' not found. Available sites: [pytorch_docs tensorflow_docs langchain_docs]

The FAILED rows in the summary are for sites that exist in the config but errored during the crawl itself.

Watch Mode

Watch mode enables scheduled periodic re-crawling of documentation sites. The scheduler tracks the last run time for each site and automatically triggers crawls when the configured interval has elapsed.

Usage

# Watch a single site with 24-hour interval
./doc-scraper watch -site pytorch_docs -interval 24h

# Watch multiple sites
./doc-scraper watch -sites pytorch_docs,tensorflow_docs -interval 12h

# Watch all configured sites weekly
./doc-scraper watch --all-sites -interval 7d

Interval Format

The interval supports standard Go duration format plus day units:

• 30m - 30 minutes
• 1h - 1 hour
• 24h - 24 hours
• 7d - 7 days
• 1d12h - 1 day and 12 hours

State Persistence

Watch mode persists state to <state_dir>/watch_state.json, tracking:

• Last run time for each site
• Success/failure status
• Pages processed
• Error messages (if any)

This allows the scheduler to resume correctly after restarts, only running sites when their interval has elapsed.

Example Output

INFO Starting watch mode for 2 sites with interval 24h0m0s
INFO Watch schedule:
INFO   pytorch_docs: last run 2024-01-15T10:30:00Z (success, 1500 pages), next run 2024-01-16T10:30:00Z
INFO   tensorflow_docs: never run, will run immediately
INFO Running crawl for 1 due sites: [tensorflow_docs]
...
INFO Next crawl: pytorch_docs in 23h45m (at 10:30:00)

Graceful Shutdown

Watch mode handles SIGINT/SIGTERM gracefully: it stops the scheduler and cancels any in-progress crawl, letting the crawler flush its BadgerDB state and partial output first, so the interrupted crawl resumes cleanly on the next run.

Run (JSON Task Spec)

The run command reads a single JSON object from stdin and dispatches the equivalent crawl or watch. It is meant for orchestration agents that would rather build a JSON payload than assemble shell flags. Unknown fields are rejected so typos surface immediately; logs go to stderr and the exit code matches the equivalent flag-driven subcommand.

{
  "command":     "crawl" | "watch",   // required
  "config":      "config.yaml",        // optional, defaults to config.yaml
  "site":        "site_key",           // exactly one of site | sites | all_sites
  "sites":       ["a", "b"],
  "all_sites":   true,
  "resume":      false,                // crawl only
  "incremental": false,                // crawl only (implies resume)
  "full":        false,                // crawl only (mutually exclusive with incremental)
  "interval":    "24h",                // watch only, defaults to 24h
  "loglevel":    "info",               // defaults to info
  "json_logs":   false,                // emit slog records as JSON on stderr
  "pprof":       ""                    // crawl only, e.g. localhost:6060
}

Examples:

echo '{"command":"crawl","site":"pytorch_docs"}' | doc-scraper run
echo '{"command":"crawl","all_sites":true,"incremental":true,"json_logs":true}' | doc-scraper run
echo '{"command":"watch","sites":["pytorch_docs","tensorflow_docs"],"interval":"6h"}' | doc-scraper run

MCP Server Mode

The crawler can run as a Model Context Protocol (MCP) server, enabling integration with AI assistants like Claude Code and Cursor.

Available MCP Tools

Usage

The MCP server uses the stdio transport, compatible with Claude Desktop, Claude Code, and Cursor.

./doc-scraper mcp-server -config config.yaml

Claude Code Integration

Add to your Claude Code configuration (claude_code_config.json):

{
  "mcpServers": {
    "doc-scraper": {
      "command": "/path/to/doc-scraper",
      "args": ["mcp-server", "-config", "/path/to/config.yaml"]
    }
  }
}

Tool Examples

List available sites:

Tool: list_sites
Result: Returns all configured sites with their domains and crawl status

Fetch a single page:

Tool: get_page
Arguments: { "url": "https://docs.example.com/guide", "content_selector": "article" }
Result: Returns page content as markdown with metadata

Start a background crawl:

Tool: crawl_site
Arguments: { "site_key": "pytorch_docs", "incremental": true }
Result: Returns job ID for tracking progress

Check crawl progress:

Tool: get_job_status
Arguments: { "job_id": "abc-123-def" }
Result: Returns status, pages processed, and completion info

Enumerate crawled pages:

Tool: list_pages
Arguments: { "site_key": "pytorch_docs", "max_results": 50, "offset": 0 }
Result: Returns up to 50 page entries (URL, title, depth, crawled_at, content_length), sorted by URL. Use offset for pagination.

Cancel a running crawl:

Tool: cancel_crawl
Arguments: { "job_id": "abc-123-def" }
Result: Returns cancelled: true/false and the job's current status. Has no effect on jobs already in a terminal state.

Contributing

Contributions are welcome! Please feel free to open an issue to discuss bugs, suggest features, or propose changes.

Pull Request Process:

1. Fork the repository
2. Create a feature branch (git checkout -b feature/amazing-feature)
3. Commit your changes (git commit -m 'Add some amazing feature')
4. Push to the branch (git push origin feature/amazing-feature)
5. Open a Pull Request

Please ensure code adheres to Go best practices and includes appropriate documentation.

Privacy Policy

doc-scraper collects nothing: no telemetry, no analytics, no accounts. All output and state stays on your machine, and the only network requests it makes are the crawls and fetches you explicitly ask for. Full policy: PRIVACY.md.

License

This project is licensed under the Apache-2.0 License.

Acknowledgements

• GoQuery for HTML parsing
• html-to-markdown for conversion
• BadgerDB for state persistence
• mcp-go for MCP server implementation
• go-readability for content extraction fallback
• modernc.org/sqlite for the pure-Go crawl-history index