How to Discover Connections in Your Reading Library with Clip

Before the knowledge graph becomes useful, you need clips. The graph draws edges between clips when their semantic similarity exceeds a threshold (around 0.6 cosine similarity by default). With only 5 clips, you will see isolated nodes floating in space. With 50 clips, real structure begins to emerge — clusters form, edges connect related ideas, and patterns become visible.

Aim for at least 20 articles on topics you genuinely care about. Use the Clip Chrome extension while browsing or save articles directly through the web app at https://api.clip.pyckle.co. Clip saves the full text content of each page, not just the URL, so the embeddings are based on actual article content rather than metadata.

The more intentional your reading, the more useful the graph. Save articles you actually read and find valuable, not everything you come across. Quality matters more than quantity here — a focused library of 30 articles on machine learning interpretability will produce a more meaningful graph than 200 random saves.

Navigate to https://api.clip.pyckle.co and click "Graph" in the top navigation bar. The graph loads your most recent 150 clips by default and runs the force simulation. The first load may take a few seconds while ChromaDB queries your library and calculates similarity scores between all pairs of clips.

You can adjust the number of clips loaded using the limit parameter in the URL (for example, ?limit=300 to load more). For very large libraries, start with a smaller limit to keep the visualization responsive, then increase it once you understand the structure.

Once the graph loads, you will see nodes scattered across the canvas. The force simulation pulls connected nodes together and pushes unconnected nodes apart, so clusters naturally form over a few seconds as the simulation settles.

Nodes are coloured by topic cluster — the same cluster labels you see in your clips list view. Articles that share a cluster and have strong semantic connections will clump together in the simulation. A tight, dense cluster indicates those articles are genuinely related at a content level, not just by topic label.

A loose, spread-out cluster tells you something different: those articles share a label but may be semantically diverse. This can happen when a topic is broad (like "technology") or when the cluster contains articles from different angles on the same theme.

Identify 2-3 prominent clusters and ask yourself: is this what I thought I was reading about? Often the graph reveals reading patterns you were not consciously aware of — a cluster of articles on productivity that you did not realize you had accumulated, or unexpected connections between business strategy and cognitive science.

Hover over any node to see the article's title and source domain in the tooltip. Click the node to open that article in the reader. Edge thickness and opacity represent similarity strength — thicker, more opaque edges indicate stronger semantic connections between clips.

Navigate from one article to its connected neighbours by going back to the graph and clicking on adjacent nodes. This edge-following often reveals surprising relationships that linear browsing would never surface. An article you saved about startup pricing might connect to one about behavioral economics you saved six months earlier. An essay on writing style might link to one about software architecture through shared concepts of clarity and structure.

These cross-temporal, cross-topic connections are the real value of the graph. Your reading accumulates over time, and the graph makes that accumulation visible and navigable.

Isolated nodes — clips with no edges above the similarity threshold — tell you something valuable. Either those articles are genuinely unique in your library (no similar content saved), or they belong to a topic cluster you have not built up yet.

Isolated nodes point to reading gaps. If you find a dense cluster of 12 articles on machine learning interpretability but 3 isolated nodes on adversarial examples, you have identified a gap worth filling. Those isolated nodes are seeds — they mark the beginning of a cluster that does not exist yet.

Pay attention to isolated nodes that interest you. They suggest directions your reading could grow. Add a few more articles on that topic and watch them connect in the graph.

The graph defaults to a minimum similarity of 0.6. Lower this threshold using the URL parameter (for example, ?min_similarity=0.5) to see more connections. This is useful for sparse libraries or when you want to explore weaker but still meaningful relationships.

Raise the threshold to 0.75 or higher for a cleaner graph that shows only the strongest semantic connections. This strips away noise and highlights the core structure of your library — the articles that are most deeply related.

There is no wrong setting. It depends on what you are trying to see. A low threshold reveals the full web of connections; a high threshold reveals the backbone. Experiment with different values as your library grows.