Sources, filtering, and clusters

Control the pipeline input and understand why an item passed or disappeared.

Supported sources

SemanticHub currently supports:

  • websites publishing RSS, Atom, or XML sitemaps;
  • public Telegram channels;
  • public Bluesky profiles.

Facebook URLs are rejected. Other social platforms are not currently supported as source types.

Adding a website

You add sources in the goal wizard, or later in Goal → Sources → Add source — by searching the library, pasting a site address, or giving a feed address directly. SemanticHub normalizes the domain, discovers feeds and sitemaps, reads basic metadata, and checks whether the source already exists. The address must resolve to a publicly routable host.

Separately, the goal's research adds portals it finds on its own; they appear under Suggested portals for you to accept or dismiss.

Reading source health

  • Active means that a goal uses the source and it is scheduled for crawling.
  • Last crawl shows when SemanticHub last attempted to check it.
  • 24-hour traffic shows recent item throughput.
  • Broken indicates that the source returns no usable content or repeatedly fails.

Do not assume a fixed time to first data. The source schedule and its own publishing cadence both matter.

Relevance filtering

By default, the model receives a title and a short excerpt. Full-content evaluation supplies more context but processes substantially more tokens.

Filter decisions and their reasoning live in the Filter tab; what the filter let through is in the Feed, behind the Before filter / After filter switch.

A useful filter instruction contains:

  1. the subject and relevant geography;
  2. event types that should pass;
  3. concrete rejection cases;
  4. exceptions where a title alone may be misleading.

Disabling filtering or leaving the instruction empty accepts every new item for grouping.

Clustering controls

SettingWhat it controls
Similarity thresholdHow close two items must be. A higher value creates narrower topics.
Minimum articlesHow many items a topic needs before it can proceed.
Immediate thresholdHow many items let processing begin without waiting for holdback.
HoldbackThe maximum collection window after a topic reaches its minimum.
DebounceA short window that batches new items before regrouping.
Article lifetimeHow long an item remains active for this goal.

Tune with evidence

  • Too many irrelevant items: add negative examples to the relevance instruction.
  • Relevant items are rejected: enable full content or broaden the instruction.
  • One event splits into many clusters: lower the similarity threshold.
  • Different events merge together: raise the similarity threshold.
  • Topics take too long to proceed: lower the minimum or shorten holdback.

All pages