How SemanticHub works

The product model, core objects, and the complete path of a publication.

SemanticHub is an automated content-intelligence pipeline. It is not merely a news-reading dashboard. Its job is to judge the relevance of incoming articles, combine the related ones into a cluster, and deliver a result where the downstream work happens.

The system relies on a handful of concepts, each defined in exactly one place — the glossary. The rest of this page assumes you know at least five of them: goal, source, article, cluster, and result.

  1. 1Fetch
  2. 2Relevance
  3. 3Embedding
  4. 4Clustering
  5. 5Workflow
  6. 6Result
  7. 7Review
  8. 8Delivery
The signal path. Filtering happens before embedding and clustering, so material rejected as irrelevant never reaches the rest of the pipeline.

The path of an item

Fetch

The crawler checks a source on its schedule, records new URLs, and extracts their content.

Evaluate relevance

The item is evaluated independently for every goal connected to the source. If no goal accepts it, processing stops.

Embed and group

Accepted material receives one shared embedding and joins a cluster based on semantic similarity and publication time.

Promote

Once the cluster meets the goal’s conditions, it becomes a result. Optional research can enrich a small cluster first.

Process

In direct mode, the cluster is the result. With a workflow enabled, AI steps produce sections defined by the goal’s output contract.

Review and deliver

The result lands in Post moderation, where it waits for approval, or it is sent automatically according to the goal’s delivery policy.

Two ways to finish the pipeline

ModeOutputUse it when
DirectCluster description and source articlesThe downstream system performs its own processing or you need monitoring without another model call
WorkflowResult sections generated by AI stepsThe result should be a draft, analysis, or other structured material

What stays under human control

  • the choice of sources and relevance filter;
  • grouping thresholds and the moment of execution;
  • the models and prompts of workflow steps;
  • the final result format — the sections a result is made of;
  • the decision to require manual approval or deliver automatically;
  • turning off test mode before a real send.

All pages