Skip to content

Graph Explorer

Cross-sector force graph of every graph-capable semantic model in dbt-cerebro. Trust relations, ownership, LP positions, validator control, bridge flows — all in one canvas.

What it is

Graph Explorer renders an ECharts force graph centered on a seed (typically an EVM address). Nodes come from the semantic registry; edges are fetched per query. Every graph-capable dbt model contributes a profile; chip toggles on the strip below the graph let you mix profiles in/out.

Resource URI: ui://cerebro/graph_explorer.

When to use it

  • "Who owns this Safe and what else do they control?"
  • "Which validators are funded by which deposit address?"
  • "Which DEX pools share LPs with this one?"
  • Sector-spanning investigations (Circles + Safe + GPay + DeFi all on one canvas).

Two screens

  1. Catalog (empty state) — lists every profile by sector. Primary action: paste an EVM address; the backend auto-detects which profiles apply via int_execution_address_roles_current.
  2. Graph screen — force-directed canvas with topbar (window, max neighbors, layout), chip strip (per-profile toggles), and side panel (node metadata, role badges, semantic provenance, suggested next hops).

Step-by-step tutorial

1. Seed from an address

open_graph_explorer()
load_graph_explorer_seed(address="0xabc…")
# Auto-detects roles, picks default profiles, renders 1-hop subgraph.

2. Expand a node

expand_graph_explorer_node(node_id="address:0xdef…")
# One additional hop, capped at MAX_HOPS = 5.

3. Update focus

update_graph_explorer_focus(
  window_days=30,
  max_neighbors=50,
  layout="circle",                 # or "force"
  status_filter="approved",        # or "candidate" / "all"
  active_profiles=["circles_trust", "lp_in_pool", "lending_user_to_reserve"],
)

Tool reference

Tool Purpose
open_graph_explorer() Open the panel (catalog screen if no seed)
load_graph_explorer_seed(address) Seed the graph on an address
expand_graph_explorer_node(node_id) One-hop expansion (cap MAX_HOPS=5)
update_graph_explorer_focus(...) Window, neighbors, layout, status filter, profiles

Profiles (initial release)

Profile Source → Target Quality
circles_trust circles_avatar → circles_avatar approved
circles_trust_history circles_avatar → circles_avatar candidate
circles_avatar_balances circles_avatar → token approved
safe_ownership address → safe candidate
gpay_ownership address → gpay_wallet candidate
token_transfers address → address candidate
lp_in_pool address → pool approved
pool_contains_token pool → token approved
lending_user_to_reserve address → token approved
validator_controlled_by address → validator approved
deposit_to_validator address → validator approved
bridge_user_flows address → bridge candidate
address_labeled_as address → project_label approved

Cross-sector relationships

These edges drive the "suggested next hops" chips in the side panel:

  • gpay_wallet_is_safe — GPay Safes are Safes.
  • ga_user_controls_gpay — canonical EOA → GP Safe binding via the Delay Module. Do NOT use safes_current_owners.owner (returns sentinel 0x…0002).
  • circles_avatar_is_address — avatars are addresses.
  • deposit_to_validator_identity — withdrawal_credentials joins GBCDeposit and consensus.
  • validator_address_is_safe — the 20-byte withdrawal address is often a Safe.
  • address_labeled_as_project — universal Dune-label enrichment.

Worked example

# Triaging a suspicious deposit address
> open_graph_explorer()
> load_graph_explorer_seed(address="0xdep…")
< 1-hop subgraph: 12 validators funded, 3 token transfers, 1 Safe ownership

> expand_graph_explorer_node(node_id="safe:0xs…")
< Safe owners: 5 EOAs

> update_graph_explorer_focus(active_profiles=["lp_in_pool", "bridge_user_flows"], window_days=30)
< Adds LP positions and bridge flows for every node in current view

# Click a node → side panel shows role badges, semantic provenance, and "next hops"

Knobs

Setting Default Where
MAX_HOPS 5 tools/graph_explorer.py
DEFAULT_WINDOW_DAYS 90 same
DEFAULT_MAX_NEIGHBORS 25 same

Best practices

  • Start with approved status filter. Candidate profiles are useful but noisier.
  • Don't expand past 3 hops without a reason — the graph becomes illegible fast.
  • Use Recenter from the side panel to reseed cleanly on a new focal node — beats expanding outward forever.
  • The Ask button pushes the current view state into model context and asks the LLM to summarise the subgraph — useful for narrative wrap-up.

Pitfalls

  • Treating safe_ownership as full-fidelity Safe lineage. Source is int_execution_safes_current_owners (current snapshot only). Use historical Safe queries for time-series analysis.
  • Mixing too many profiles. Three to five active profiles is usually the legibility sweet spot.
  • Following gpay_ownership instead of ga_user_controls_gpay. The former is the Safe-owner relationship (sentinel-poisoned for GPay Safes); the latter is the canonical EOA-to-GP-Safe binding.

See also