Who appears with whom? The entity graph on the Topic Map

Until now the Topic Map showed one thing: the concepts of your wiki and how they connect. We wrote about that in Your reading, as one knowledge graph. That view answers the question of what you have been reading about.

A second question comes up in consulting work at least as often: who and what appeared in it? Which vendors show up in your articles on security incidents, which authorities in the ones on reporting duties, which products next to which regulations? For that the Topic Map now has a switch: Concepts or Entities.

What the entity graph shows

Pith is a reading memory for consultants: from the articles you save it writes a cited wiki and a knowledge graph that shows which people, companies, products and regulations appear together in your sources. Every time you save something, Pith recognises the names in the article: people, companies and authorities, products, regulations and named methods. The entity graph turns them into a network.

• Nodes are the names. Their colour shows the type, their size how many of your bookmarks mention them.
• Links appear when two names are mentioned in the same bookmarks.
• A click on a node opens a side panel: the bookmarks the name appears in, the names it appears with, and a link to the entity's page.

For example: over a few weeks you saved articles on cloud sovereignty. The graph then shows a cloud provider next to a regulation such as the Data Act and an authority, because they appeared in the same texts. Which articles those were is one click away.

Why ubiquitous names don't tie everything together

A naive graph simply counts how often two names occur together. In an IT context the result is predictable: GitHub and OpenAI sit in the middle, connected to everything. They occur everywhere, so they also occur together with everything. That says nothing about your subjects.

So the graph measures the strength of a link relative to how common each of the two names is overall. Two names that almost only appear together are strongly linked. A name that is in every other article gets only a weak link to any single partner. For each entity the graph shows its six strongest links.

That keeps visible the names that actually belong together — and the ones that are everywhere become what they are: background.

Why cleaning matters more than drawing

Drawing a graph is the easy part. Whether it is any good is decided earlier, by what becomes a node in the first place. When we first looked at the existing entities as a graph, three problems stood out.

Generic nouns had become entities. "Secrets", "ConfigMaps", "personal data servers" — terms, not names. In the graph they would have attached themselves to everything involving Kubernetes or data protection. Extraction now takes proper nouns only, and when the existing entities were cleaned up, such terms were archived. In our own test workspace that was 20 of 69 entries.

Spelling variants were separate entities. "Postgres" and "PostgreSQL" stood as two nodes side by side. Now the model gives the official name as the name and short forms as aliases, and it receives the names that already exist in your workspace, with the instruction to reuse them. On top of that there is a match on the spelling: if one name is the start of the other, with at least five shared and at most three extra characters, it is the same entity. "Postgres" and "PostgreSQL" are merged, "Java" and "JavaScript" are not. Merges are reversible.

The types did not fit. There were only person, company and framework. PostgreSQL, Redis and Kafka were filed as companies; NIS2 and the AI Act would have ended up as frameworks. Now there are two more types: product for software, tools, platforms and models, and regulation for laws, directives, standards and certifications. PostgreSQL is a product, NIS2 a regulation, and framework now only stands for named methods such as MECE.

The clean-up only touches entities that were recognised automatically. Anything you created yourself stays as it is.

Archiving as a block list

One detail only showed up in the real run, and it stayed because it does exactly the right thing. An archived term remains known in your workspace. If the model recognises "Risk management" again in a later article, the mention is attached to the archived entry — and the graph and the lists keep hiding it.

So archiving does not just mean "get rid of it", but "this is not a name here, next time either". Every generic term you archive keeps the graph cleaner for good.

Honest about the limits

The entity graph grows with your bookmarks, and it cannot show more than you saved.

By default only names that appear in at least two bookmarks are shown. A name that turns up once is not a pattern yet. With few bookmarks, though, hardly anything remains. In our own workspace with seven bookmarks, Pith recognised 33 entities, but only two of them appeared in more than one bookmark. That is why the page lowers the threshold to one bookmark once, automatically, when fewer than five names would otherwise be visible. You can also change it yourself.

A sparse graph is not a bug but an honest state. Every saved article adds names and links, and the links that appear then have a reason: there are texts in which both names stand, and you can open them.

What about Raindrop, mymind or Recall?

Many bookmark managers organise with AI, and some do it very well. The question is what gets organised: the filing or the content.

• Raindrop.io offers collections, nested folders and tags, strong full-text search and AI suggestions for tags and collections on save. That keeps even a large collection tidy. Its bookmarks stay separate, by design.
• mymind tags and groups everything automatically, with no folders at all, and is especially pleasant for visual saving. Each card stands on its own; there are no citations across sources.
• Recall comes closest to a graph: it summarises saved articles, videos and podcasts and links related items in a knowledge graph.

Tags and collections answer "where did I put it?". A graph that links related items answers "what is connected to what?". The entity graph goes one level deeper, to the names inside the texts: it links a regulation to the vendors and authorities that appear in the same articles, and takes you to exactly those articles with one click. There is no filing to maintain.

The one-to-one comparisons: Pith vs Raindrop, Pith vs mymind and Pith vs Recall. An overview of bookmark managers with AI orders the tools by how deep their AI goes.

When which view helps

The two views complement each other. The concepts show what your reading added up to: subjects your wiki has written pages about. The entities show who and what appears in it. Before a meeting with a client in the energy sector, one question might be "What do I know about supply-chain security?" and the other "Which vendors and regulations appear together in my articles about it?"

Both answers lead back to the same sources — the articles you saved yourself.