Skip to content

BETA.md

Latest commit

 

History

History
73 lines (52 loc) · 2.8 KB

BETA.md

File metadata and controls

73 lines (52 loc) · 2.8 KB
Raw

Graphsight: Beta for Friends

Thanks for trying this. I want your honest take, including "this is cool but I wouldn't use it."

What it actually is

Graphsight captures exactly why your AI agent retrieved what it did.

It records every retriever call, per-item scores, which retrieval arm produced each result (vector vs graph), and the relational paths between results (e.g. person → PR → Service → Issue), then renders it as an interactive graph in the browser: every node clickable, with the underlying evidence, score, and source link in an inspector, plus the execution timeline of the run that produced it.

You get full visibility into the retrieval reasoning instead of a black box.

Two pip packages. No backend, no account, nothing leaves your machine.

Try it on your own repo (5 minutes)

pip install graphsight "graphsight-langgraph[example]"

# ingest + trace a query against your GitHub repo
graphsight-github-trace yourusername/yourrepo "who changed authentication recently?"

# open the graph in your browser
graphsight graphsight_out/trace_state.json

Now the graph is your PRs, your issues, your commits, your people. what matched the question, who authored it, which issues it resolves.

Works on public repos with no token. Private repos need GITHUB_TOKEN. Repos with no PRs or issues still work, recent commits carry the history.

Tip: ask who / what / when questions ("who touched auth recently?"), not how-does-it-work questions, commit history describes changes, not architecture, and the tool won't pretend otherwise.

If you already build with LangGraph, wiring the tracer into your own agent is one callback handler:

tracer = LangGraphTracer()
graph.invoke(inputs, config={"callbacks": [tracer]})

See graphsight-langgraph/README.md.

What I want your blunt feedback on

  1. Did the graph tell you anything useful that the raw answer didn't? If not, this is decoration and I need to know.
  2. Would you actually wire this tracer into a real agent you run regularly? Why / why not, trust, effort, or no need?
  3. What's missing before this becomes part of your workflow? Comparing two runs? Auto-capture on failure? Better filtering? Something else entirely?

Drop feedback as a GitHub issue on this repo, or DM me directly.

Current limitations (no surprises)

  • The GitHub trace uses simple lexical matching + 1-hop graph expansion for now, labeled as such in the output, and sometimes dumb.
  • Scores come directly from your retriever, nothing is invented. Missing scores render as missing.
  • Sync and async LangGraph runs are both verified by the test suite.
  • The hosted Graphsight engine (live graph memory over MCP) is still in development, this beta is the local viewer + tracer, which need nothing.