Smart Connections

Why Connections Matter More Than Content

A single note about "microservices vs monoliths" is useful. But that same note connected to a bookmark on "scaling team autonomy," meeting notes about "service ownership," and an article on "Conway's Law" — that's an *insight*. The individual pieces are trivia. The connections are knowledge.

The best second brain systems don't just store — they link.

Finding Related Items

The simplest connection: for any item, find the N most similar items using the embeddings you already have.

function findRelated(targetId: string, allChunks: Chunk[], topN: number = 5): Chunk[] {
  const target = allChunks.find(c => c.id === targetId);
  if (!target) return [];

  return allChunks
    .filter(c => c.id !== targetId)
    .map(c => ({ chunk: c, score: cosineSimilarity(target.embedding, c.embedding) }))
    .sort((a, b) => b.score - a.score)
    .slice(0, topN)
    .map(r => r.chunk);
}

This gives you "items like this one" for any piece of knowledge. Useful, but not surprising — similar items tend to be obviously related.

Cross-Source Connections

The highest-value connections span source types:

To find cross-source connections, add a filter: only return results from a *different* source type than the input.

Building the Knowledge Graph

A knowledge graph makes connections explicit and navigable:

interface GraphEdge {
  sourceId: string;
  targetId: string;
  score: number;
  crossSource: boolean;  // Different source types?
}

function buildGraph(chunks: Chunk[], threshold: number = 0.75): GraphEdge[] {
  const edges: GraphEdge[] = [];
  for (let i = 0; i < chunks.length; i++) {
    for (let j = i + 1; j < chunks.length; j++) {
      const score = cosineSimilarity(chunks[i].embedding, chunks[j].embedding);
      if (score >= threshold) {
        edges.push({
          sourceId: chunks[i].id,
          targetId: chunks[j].id,
          score,
          crossSource: chunks[i].source !== chunks[j].source,
        });
      }
    }
  }
  return edges;
}

Tuning the Threshold

The similarity threshold controls the density of your graph:

Start at 0.75 and adjust based on your data. If the graph is too sparse, lower it. If every item connects to everything, raise it.

Surfacing Surprising Connections

The most valuable connections are surprising — items that are semantically similar but differ in source type, topic area, or time period.

A "surprise score" combines semantic similarity with contextual distance:

surprise = similarity * (1 + source_bonus + time_bonus)

Where source_bonus adds 0.2 if the sources differ, and time_bonus adds 0.1-0.3 based on how far apart the dates are. A connection between a note from January and a bookmark from June is more surprising (and often more insightful) than two notes from the same day.

Clusters and Themes

Beyond pairwise connections, clusters reveal themes in your knowledge:

Key Takeaways