Void Search

A fork of SearXNG, the metasearch engine. I run this as my own instance with the defaults I actually want and a few parts SearXNG does not have. All credit for the engine itself goes to the SearXNG project.

What is different

Most of the changes are defaults, not new features:

• public_instance and the limiter are on. Valkey is required.
• POST searches only. No access logs, no query logs.
• Engines never see your IP or browser: outgoing requests get a random browser user-agent and a generic header set.
• Results are cached for 180 seconds so repeat searches do not re-query engines.
• English by default. Results in other scripts are dropped on en queries.
• A host blocklist plus demotion for SEO-farm domains.
• Engines that ban or CAPTCHA instance IPs (Google, DuckDuckGo, Brave, Startpage, Qwant, Mojeek and friends) are force-disabled by the engine policy. Set void.engine_policy: false or SEARXNG_ENGINE_POLICY=false if you want them anyway.
• /stats, download formats, timings, and version strings are not exposed.
• On Linux, the process drops into a Landlock sandbox after startup. Skipped in debug mode.

Performance defaults: four Granian workers, an HTTP/2 pool of 256 connections, max eight redirects, two second request timeout.

The WAF in front is RavenGuard. The Landlock bindings are landlockpy.

Run locally

Python 3.12 or newer:

python3 -m venv .venv
.venv/bin/pip install -r requirements.txt -r requirements-server.txt
export SEARXNG_SECRET="$(openssl rand -hex 32)"
export SEARXNG_DEBUG=1
export SEARXNG_PORT=48731
export SEARXNG_BIND_ADDRESS=0.0.0.0
export SEARXNG_LIMITER=false
export SEARXNG_PUBLIC_INSTANCE=false
PYTHONPATH=. .venv/bin/python -m searx.webapp

The limiter and public_instance need Valkey. For a single-user run without it, leave both false. Debug mode skips Landlock so the reloader works.

Docker with RavenGuard

mkdir -p config
export RG_CHALLENGE_SECRET="$(openssl rand -hex 24)"
docker compose up -d

client -> RavenGuard :48731 -> Void Search :8080 -> engines

Void itself is not published on the host. Change RG_CHALLENGE_SECRET before binding anything publicly. Raise GRANIAN_WORKERS if the box has RAM to spare.

Coolify

docker-compose.coolify.yml is the Coolify-ready definition: RavenGuard in front, Void Search behind it, and a Valkey sidecar for the limiter and result cache.

In Coolify create an Application from this repository, select the Docker Compose build pack and set the Compose location to docker-compose.coolify.yml. Put the public domain on the ravenguard service's Domains field with the container port it exposes, for example https://search.example.com:8080. Set RG_CHALLENGE_SECRET and any SEARXNG_* or RG_* overrides in Coolify's environment editor; the secret key is generated on first boot and kept in the void-config volume.

The image lives at ghcr.io/quad4-software/void-search. If the package is private, add a registry credential in Coolify or mark the package public.

Wiby and YaCy

Both ship disabled. In /etc/searxng/settings.yml:

engines:
  - name: wiby
    disabled: false
  - name: yacy
    disabled: false
    base_url:
      - https://yacy.searchlab.eu

If you have your own YaCy, point at that. Public peers rate-limit quickly.

Blocklist

Edit searx/void_blocklist.txt or set void.block_hosts in settings. Matching hosts are removed from results.

The crawler

crawler/ is a Rust crawler that builds our own index. The thevoid engine reads it. It respects robots.txt including crawl-delay, keeps one request in flight per host, backs off on 429 and 5xx, refuses private IPs, skips blocklisted hosts, detects tarpits (slow-drip responses get aborted), and solves Anubis proof-of-work natively (codeberg, freedesktop and friends). Cloudflare-type JS challenges go through FlareSolverr when VC_FLARESOLVERR is set, and outbound traffic can ride a socks5 proxy (Mullvad's in-tunnel proxy at socks5h://10.64.0.1:1080 via VC_PROXY). The process sandboxes itself with Landlock when the kernel supports it.

Documents are zstd-compressed in redb and indexed by tantivy. Ranking is BM25 on title and body with an authority boost from inlinks and a freshness decay.

cd crawler
cargo build --release
./target/release/void-crawler crawl --seeds-file seeds.txt
./target/release/void-crawler serve   # search api on 127.0.0.1:8088

seeds.txt ships ~1400 quality seeds: wikipedia topic pages, the curated kagi small web domains and feeds, and community/reference sites. RSS and Atom feeds in the seed list are mined for links rather than indexed.

Or run does both in one process. In compose, the crawler service is profile-gated: docker compose --profile crawler up. In Coolify it builds with docker-compose.coolify.yml automatically.

export dumps the document store - --format jsonl for our pipe format or --format hister for the exact bracketed layout hister import file reads. import accepts both shapes, so a hister export can seed or merge into our index and vice versa.

Live hister integration: set VC_HISTER_URL (+ VC_HISTER_TOKEN) and every indexed page is pushed to hister through /api/batch as it is crawled. void-crawler push-hister backfills an existing store when the server was down or push was off. And searxng can search a hister index directly - the hister engine (HISTER_API_URL, HISTER_TOKEN, off by default) merges your personal index into results.

Settings

Setting	Void default
server.public_instance	true
server.limiter	true
server.method	POST
search.default_lang	en
void.anonymize_outgoing	true
void.drop_foreign_script	true
void.landlock	true
void.no_logs	true
void.engine_policy	true
void.result_cache_ttl	180
outgoing.request_timeout	2.0s
outgoing.pool_connections	256
outgoing.max_redirects	8

Override in /etc/searxng/settings.yml with use_default_settings: true.

License

AGPL-3.0-or-later, same as SearXNG. See LICENSE and NOTICE.