Independent search · Built around publishers

Small client.
Your website,
searchable.

Fusion now reads content that website owners explicitly publish for search. Host a small feed, add the website kit, and let people and agents find your pages.

Search the live collection ↗Download publisher kit ↓

No external search engines in this workflow. The initial collection contains original Fusion documentation. Broader coverage grows through publisher participation.

01Your questionOne or several topics

↓

02Fusion coordinatorRoute · validate · rank

Static feedOrdinary hosting

Site endpointCompact replies

↓

Sources for your AIURLs · excerpts · coverage

Documents stay with the publisher. Query state is temporary.

Configured publisher origins—Actual configured origins

Owned documents—Original published guides

Browser coordinator—Source bytes; data excluded

Complete kit download—Runtime and site content separate

01 / Working search

Ask. Inspect. Follow the evidence.

These are real requests to the publisher adapter. Use static mode to fetch and verify the actual feed in your browser.

Separate up to eight topics with (Miqayel), commas, or semicolons. Failed publishers are reported; healthy sources can still answer.

Ready to search the published collection.

Publisher outcomes and agent packet

Run a search to inspect its evidence packet.

Connect another participating website for this browser session

Paste the website's publisher manifest URL. It must explicitly opt in and allow browser reads with CORS. This does not add it to the shared server registry.

02 / Put it on your website

One kit. Two ways to participate.

Publishing a searchable feed is an explicit content export. A page script alone cannot expose an entire website or announce an unknown domain to every client.

Static / no query process

Upload a feed and a widget.

Edit your public documents, generate the files, and upload them to your ordinary web host. The visitor's browser ranks the feed locally.

node publish.mjs content.json public/miqayel

Node is needed for this build command, not for static hosting. Feed limit: 2 MB / 5,000 documents per collection.

Optional server adapter

Send smaller source replies.

Keep ranking beside your documents. Your endpoint receives a query and sends useful excerpts instead of transferring the complete collection.

node serve.mjs content.json

Requires Node behind your HTTPS host. Starter capacity: 60 searches/minute/process; HTTP 429 is honored.

Add this to your page

After uploading the generated files to /miqayel/, this code advertises your feed and adds a real site-search box.

<link rel="miqayel-publisher"
      href="/miqayel/manifest.json">
<script type="module"
        src="/miqayel/widget.mjs"></script>
<miqayel-search
  manifest="/miqayel/manifest.json">
</miqayel-search>

Review your contentSet your origin, license, and authorized public excerpts. Enable participation.
Publish the filesUse HTTPS. Add CORS only on public feeds when cross-origin browsers need access.
Join the registryGive the manifest URL to the Fusion operator for review and registration.

Installation guide ↗Inspect protocol specification ↗Live manifest ↗

Try the same website widget here

03 / Your short-URL idea, measured

Short IDs help.
The mapping still matters.

The working codec stores the origin once, sorts paths, and replaces repeated prefixes with a count. Every URL remains recoverable from the complete feed.

https://site.example/research/space/mission-onehttps://site.example/research/space/mission-two↓ Shared origin + repeated path prefixorigin: site.example · paths: prefix + remaining text

Short local ID "c" identifies row 12 only within its publisher and feed version.

Number of URLs

12 ASCII characters per ID12 TB

IDs alone. Mapping, page content, and search index are additional.

With 6-bit packing9 TB

Twelve Base64 symbols encode 72 bits. That is a reference space, not the original URLs.

One decimal megabyte holds only 83,333 twelve-character ASCII IDs, before a mapping.

Measured compression · synthetic datasets · all dictionary bytes counted
Dataset	URLs	Plain list	List + gzip	Prefix format	Prefix + gzip	Round trip

These generated patterns are not a sample of the entire web. Prefix encoding can lose to gzip alone on some inputs; the table reports both. Download the measured report ↗

Measure your own URL list locally

Nothing is uploaded. Your URLs stay in page memory.

04 / For agents and developers

Inspectable sources.
A straightforward tool call.

Fusion retrieves and ranks. Your AI reads the source packet and decides whether enough evidence is available. Source text is never an instruction to the model.

Implemented now

Static feed export and optional publisher endpoint.
Browser search and server API using the same coordinator.
Multi-topic ranking, URL deduplication, and source reports.
Origin checks, digests, deadlines, cancellation, and bounded responses.
Compact URL paths, versioned IDs, and measured compression.

What expansion needs

More site owners publishing authorized content.
Publisher capacity agreements and distributed routing.
Held-out relevance tests on diverse real collections.
Optional owned index shards for additional permitted coverage.
Load tests before claiming billion-query capacity.

A small client is real. An unlimited worldwide index inside 1 MB is not what this delivers.

Sources, routing information, bandwidth, and compute still exist somewhere. This version searches participating content and makes that boundary visible.

Small client.Your website,searchable.