Blitz.md

Blazing-fast URL -> Markdown converter written in Rust.

Fetch any webpage and get clean, readable Markdown in milliseconds. Streams the HTML, parses and converts on the fly - no intermediate DOM, no bloat.

blitzmd https://example.com > output.md

---

Features

• Streaming pipeline - converts HTML chunks as they arrive, no wait for full download
• Zero-copy post-processing - bulk byte-slice operations, single pass for tag stripping + entity decoding
• Absolute URL resolution - all links and images resolved to full URLs
• Block link flattening - <a> used as card containers become readable text, not broken multi-line links
• Full HTML entity support - named (&,  , —...) and numeric ( , —...)
• Smart noise removal - strips <script>, <style>, <nav>, <footer>, <svg>, <iframe> and form elements
• Brotli + gzip compression - fastest possible transfer
• --verbose timing mode - inspect each pipeline stage

---

Installation

Pre-built binaries

Download the latest release for your platform from the Releases page.

Build from source

git clone https://github.com/Athenox14/Blitz.md
cd Blitz.md
cargo build --release
# binary: ./target/release/blitzmd

Requires Rust 1.75+.

---

Usage

blitzmd [--verbose] <URL>
blitzmd search [--limit N] [--timeout-ms MS] [--json] [--llm] [--deep] [--deep-limit N] [--verbose] <query>

Ultra-fast web search

# Fast metasearch (parallel providers + dedup)
blitzmd search "rust async runtime"

# LLM-ready output + telemetry
blitzmd search --llm --verbose "Athenox Development"

# JSON output for scripts
blitzmd search --json --limit 10 "vector database benchmark"

# Enrich top results by converting pages to markdown excerpts
blitzmd search --llm --verbose --deep --deep-limit 3 "tokio vs async-std"

Search behavior and concurrency

• DuckDuckGo and Bing RSS are queried in parallel.
• Results are deduplicated and merged.
• Fast-path can stop waiting early when confidence is high.
• Relevance filtering removes off-topic noise.
• Automatic rescue strategies run when providers are unstable:
• wider-timeout retry,
• targeted query rescue,
• domain discovery fallback.

Search speed tips

• Lowest latency: avoid --deep, lower --limit, reduce --timeout-ms.
• Highest quality under network jitter: keep default timeout or increase (1500-2500ms).
• --deep adds parallel page fetch/conversion cost but improves validation quality.

Verbose search telemetry

--verbose for search prints:

• provider status (ok, empty, timeout, error, skipped-fastpath),
• provider timings,
• merge timing,
• search throughput,
• deep throughput and per-page averages when --deep is enabled.

Basic conversion

# Basic conversion
blitzmd https://news.ycombinator.com

# Save to file
blitzmd https://en.wikipedia.org/wiki/Rust_(programming_language) > rust.md

# Show timing breakdown
blitzmd --verbose https://example.com

Verbose output

[timing] connect+headers : 352ms
[timing] body+parse      : 2.60ms
[timing] raw size        : 7 KB
[timing] post_process    : 24us
[timing] total           : 355ms

Real-world search performance (sample)

Measured from user runs on Windows (network-dependent):

Command	Typical observed total
search --llm --verbose	~180 ms to ~900 ms
search --llm --verbose with retries	~1.7 s to ~2.4 s
search --llm --verbose --deep --deep-limit 3	~640 ms to ~1.0 s

---

How it works

URL
 |
 v
reqwest (gzip/brotli, tcp_nodelay)
 |  streaming chunks
 v
lol_html (streaming HTML rewriter)
 |  - removes noise tags entirely (script, style, nav, footer...)
 |  - injects markdown markers around content tags (h1-># , strong->**, a->[]()...)
 |  - resolves relative URLs on the fly
 v
strip_and_decode()          <- single pass, bulk byte-slice copy
 |  - strips residual HTML tags
 |  - decodes HTML entities
 v
fix_block_links()           <- flattens multi-line [text](url) into plain text
 v
normalize_lines()           <- trim + collapse blank lines
 v
stdout

The entire post-processing pipeline runs in ~25-800us depending on page size.

---

Benchmarks

All benchmarks run on the same machine, conversion only (network excluded), against real pages.

Conversion speed - same HTML input

Page	Size	Blitz.md	html2text	trafilatura
Wikipedia - Rust lang	207 KB	813 us	169 ms	577 ms
Le Monde	82 KB	315 us	65 ms	71 ms
GitHub - linux/linux	30 KB	116 us	45 ms	128 ms
Hacker News	14 KB	65 us	14 ms	46 ms
Python docs	8 KB	55 us	9 ms	18 ms

Blitz.md is 165x to 1100x faster than Python-based alternatives on conversion.

End-to-end (fetch + convert) vs Jina AI Reader

Page	Blitz.md	Jina AI
oxalisheberg.fr	338 ms	549 ms
Python docs	339 ms	439 ms
Hacker News	660 ms	776 ms

Jina edges out on latency for some URLs thanks to distributed infrastructure and caching. Blitz.md runs entirely on your machine - no API key, no rate limits, no data leaving your network.

---

Comparison

	Blitz.md	html2text	trafilatura	Jina AI Reader
Language	Rust	Python	Python	Cloud API
Conversion speed	~25-800 us	9-170 ms	18-580 ms	N/A (remote)
Streaming	Yes	No	No	No
Local / offline	Yes	Yes	Yes	No
Private URLs	Yes	Yes	Yes	No
Rate limits	None	None	None	Yes
API key required	No	No	No	Yes (paid tiers)
Binary size	~5 MB	-	-	-

---

Development

cargo test

If your workspace blocks writes in target/ (common in synced folders):

$env:CARGO_TARGET_DIR="$env:TEMP\\blitzmd-target"
cargo test

---

License

MIT