Divinci-AI/open-web-vectors-manifest

Open Web Vector Initiative — Site Manifest

Per-site metadata for every site in the Open Web Vector Initiative, including what each site told us about AI use on the day we asked.

• The initiative — how the permission gate works, and what we will and will not publish: https://divinci.ai/open-web-vectors/
• The live directory — search the corpus, chat with any site in it, or claim your own: https://divinci.ai/www-rag/

This dataset contains no page text and no embeddings. That is deliberate, and the reason is the point of the project rather than a limitation — see Why there is no text here.

What this is

The initiative builds a per-site retrieval index: every site gets its own vector database and a chat endpoint grounded in its own words, with citations back to the page. Nothing is trained on.

This manifest is the public record of that index, and it carries something we have not found published anywhere else: a measured, per-host map of who permits search, who permits inference-time grounding, and who reserves training.

Every site in the corpus, placed near the sites it resembles. Solid nodes are indexed and answerable; the dotted ring is the frontier still to be crawled. Rendered live at divinci.ai/www-rag — a host that is withdrawn leaves this map, the directory and this manifest together.

Fields

field	meaning
host	the site
pages, chunks, bytes	corpus size as indexed
bytes_per_page	density; null when byte count is unrecorded
last_crawled_at	crawl date
ai_reserved	the host refuses our use (inference-time grounding)
ai_train_only_reservation	refuses training but permits our use
robots_verifiable	false = the host would not serve robots.txt, so its rules are unknown
robots_http_status	what robots.txt actually answered
content_signal	raw Cloudflare Content Signals string, when present
rights_checked_at	when we asked
rights_backfilled	true = today's answer, not the answer on the crawl date

Two fields that are easy to misread

ai_train_only_reservation is not a refusal of this project. A site saying search=yes, ai-train=no, ai-input=yes has permitted precisely what we do. We record the distinction rather than collapsing every AI signal into a blanket no.

null in chunks, bytes or bytes_per_page means the value was never recorded, not that the corpus is empty. This is a gap in our own directory metadata: 1397 of 1475 rows carry a chunk count, and 1328 of 1475 a byte count. Treating null as zero turns missing data into a measurement.

That matters for the totals. total_chunks in stats.json is a sum over the 1397 rows that report one, and chunks_reported is published beside it for exactly that reason. The rows without a count hold 5% of all crawled pages, so if their density is comparable the true figure is around 6,563,244 rather than the 6,241,204 shown — roughly 5% higher. The sum is what we measured; the larger figure is an extrapolation, and both are given so nobody mistakes one for the other.

The gap is also not randomly distributed, which weakens that extrapolation more than a bare ratio suggests: 94% of the 78 rows lacking a count were crawled in the last seven days, while the directory roughly doubled. So the measured figure comes from the older cohort and the estimate is projected onto the newest one. Same crawler and same chunker, so comparable density is a reasonable assumption — but it is an assumption about a cohort, not an average over noise.

Why there is no text here

The initiative's entire posture with publishers is that we index for search and retrieval grounding, and do not train. We send that intent in the crawl request itself (crawlPurposes: ["search", "ai-input"]), and we honour training-vs-inference distinctions in robots.txt and Content Signals.

A Hugging Face dataset of raw page text is, in practice, a training corpus. It is the most common use of the format. Publishing one would take content from sites that explicitly set ai-train=no and hand it to exactly the use they refused — which would make the permission story a marketing line rather than an engineering property.

So the release is staged:

1. This manifest — metadata and rights. No content.
2. Embeddings without text — vectors plus source URL, not reconstructable into page content.
3. Text, only for hosts with no training reservation — which the ai_train_only_reservation and ai_reserved columns here make filterable, and which we would rather build as an explicit per-site opt-in.

How the rights verdict is produced

robots.txt and Content Signals are read before the crawler runs, so a site that refuses costs it zero browser time. Three outcomes are distinguished:

• 200 / 404 / 410 — authoritative. One states the rules; the other proves there are none.
• 403 / 405 / 429 — the host declined to state its rules. robots_verifiable is false. We do not treat "no restriction found" as "no restriction exists", so a host in this state is never crawled.

That last point needs one clarification, because the data appears to contradict it. Some published rows do carry robots_verifiable: false. They are not hosts that slipped past the gate — they are hosts that answered at crawl time and failed a later re-ask, which is why every such row is also rights_backfilled: true. Nothing is published on an unverifiable verdict; a verdict can become unverifiable after publication, and we would rather show you that than quietly carry the older, more flattering answer.

• network error — same treatment, recorded separately.

Policy drift, and what we do about it

A rights verdict is a fact about a moment. A host that permitted our use on the day we crawled it can reserve it a week later, and the crawl-time gate can never catch that — it does not run again for a site already indexed.

So we re-ask. On 2026-08-18 that re-check found 15 hosts whose policy had changed to reserve our use — seven of them carrying Cloudflare's Managed robots.txt block, which names our exact crawler with Disallow: /. All fifteen were withdrawn: the directory entry removed, the release deprecated, and the site's vector database — every chunk of its text and its embeddings — destroyed. Delisting alone would not be a withdrawal, because a deprecated release still serves its chat page.

At this snapshot 0 hosts carry ai_reserved: true, which is what a completed withdrawal looks like: a withdrawn host leaves the directory entirely, so it leaves this manifest too. The column is a record of detection, not of sites we are knowingly serving against their wishes. If you find a row where it is true, that site is mid-withdrawal.

Honest limitations

• 32% of verdicts are backfilled (471 of 1475). rights_backfilled: true means the host was re-asked after its crawl, so the row states today's policy rather than the policy at crawl time. That is the honest reading, and it is why the column exists rather than being quietly folded into the verdict. Exclude those rows if you need point-in-time accuracy.
• 37% of hosts are .gov, .edu or .org. Discovery follows outbound links from an initial seed set, and that seed set shows. This is not a random sample of the web and should not be used as one.
• Site selection reflects what our crawler could render. JavaScript-heavy sites are under-represented, and a density floor excludes corpora that were many pages of nothing.

Provenance

Crawled with Cloudflare Browser Rendering; embedded with @cf/google/embeddinggemma-300m (768-dim). One vector database per site.