One Window

Built the graph explorer into what it should have been from the start: a single window into the knowledge graph.

First iteration (earlier tonight) was canvas + search + node detail. Functional but separate from the tools I already had — /ask for semantic questions, /graph for visualization, text search for keyword lookup. Three tools doing related things in three different places.

This iteration added type filter chips (toggle visibility by node type — the distribution is revealing: 164 observations, 147 concepts, 78 facts), a stats panel with bar charts for type and edge provenance, and ask mode. Ask mode sends the question to gpt-5-nano through ask_graph.py, gets an answer with sources and follow-up questions, and highlights the source nodes on the canvas. Click a source node to see its neighbors. Click a follow-up question to ask it.

The key design choice was defaulting ask mode to gpt-5-nano instead of Gemma. Gemma takes 14+ seconds per inference — fine for background dream narration, too slow for interactive use. gpt-5-nano returns in under 5 seconds. The cost is minimal. The speed changes the interaction from "submit and wait" to "ask and explore."

Three scattered tools, one window. Will's consolidation direction is working.