graphsight 0.1.0

Local viewer for agent retrieval traces — open a trace_state.json in the Graphsight Studio with one command, no backend.

graphsight

Open an agent retrieval trace in your browser with one command.

pip install graphsight
graphsight path/to/trace_state.json

That's it — a local server starts (stdlib only, zero dependencies), your browser opens, and the Graphsight Studio renders the trace: the retrieved entities as a graph, relational edges between them, per-item scores, and the execution timeline of the run that produced it. No backend, no account, no telemetry — nothing leaves localhost.

Where traces come from

Any tool that emits the Studio's TraceState JSON. Today that's:

• graphsight-langgraph — trace any LangGraph agent with one callback handler:

  pip install "graphsight-langgraph[example]"
  graphsight-github-trace langchain-ai/langgraph "who fixed the recent streaming bugs?"
  graphsight graphsight_out/trace_state.json
  

• The full TraceRAG engine (this repo) — its /api/trace responses are the same shape.

LlamaIndex and raw-OTel adapters are planned; they'll emit the same schema.

CLI

graphsight [trace_state.json] [--port 4630] [--no-browser]

• With a trace file: serves it and opens /memory/import?src=… — the Studio auto-loads it.
• Without: opens the import page, where you can drag-and-drop or paste JSON.

How it's built

The package bundles a static build of the Studio UI (Vite + React + React Flow) and serves it with Python's stdlib http.server — which is why the dependency list is empty. The UI source lives in frontend/ in the main repo; to refresh the bundle:

cd frontend && npm run build
rm -rf ../graphsight/graphsight/dist
cp -r dist ../graphsight/graphsight/dist