[Atlas] Cross-artifact relationship-graph view (A cites B debates C) done

Effort: thorough

Goal

The 688K-edge knowledge graph is exposed at /graph (site/graph.html)
but only over canonical entities (genes, diseases). The richer artifact_links table (which captures relationships like cites, synthesizes, derived_from, debates, supersedes, responds_to) is invisible to users.

Build a "relationship graph" viewer rooted at any artifact id that
expands its full N-hop neighborhood across artifact_links, so users
can trace "this hypothesis was synthesized by that brief, which cites
this paper, which is debated in this session, which answered this
open question". Today there's no way to see this lineage.

Acceptance Criteria

Optional query params: depth (default 2, max 4),
link_types (csv filter, default all), direction
(out / in / both, default both). {nodes: [{id, type, title, score, ...}], edges: [{src, dst,
type, created_at}]}. Bounded: hard cap of 200 nodes per
response, BFS truncates with a truncated_at_node_count flag. site/relationship-graph.html, link-type colored, node-shape
by artifact_type. Click on node = navigate to that artifact's
page. Hover = tooltip with title + composite_score + 1-line
summary. (BFS) between two artifacts as a highlighted subgraph. view (root + depth + filters) as a relationship_view artifact
so it can be cited in briefs and embedded in dashboards. asserts depth=2 returns the right neighborhood, depth=4
returns the right truncated set, and shortest-path returns the
expected 3-edge path. /relationship-graph/<top-AD-hypothesis-id>?depth=3 and
verify ≥10 nodes appear with at least 3 distinct link types
represented.

Approach

Audit artifact_links schema and the existing link types in use

(SELECT DISTINCT link_type FROM artifact_links).

BFS implementation in

scidex/atlas/relationship_graph.py — recursive CTE in PG with
depth limit + node-count cap; return ordered by hop distance.

D3 force-directed view in site/relationship-graph.html reusing

the styling from site/graph.html. New file, no edit to
site/graph.html.

Saved-view artifact: register via artifact_registry with

metadata {root, depth, link_types, direction, snapshot_node_ids}.

Path-find: standard bidirectional BFS via second recursive CTE.

Dependencies

• e352460b-2d76 — view_spec_json DSL (only used for embedded

rendering of saved views).

• Existing artifact_links table.

Work Log

2026-04-27 — Implementation complete [task:66bdea91-0996-4c4f-95d1-080373aef901]

Approach taken:

• Audited artifact_links: 20 valid link types in chk_link_type constraint; 3.4M rows dominated by see_also, related, mentions. Notable semantic types: cites (224), derives_from (11K), supports (1.5K).
• BFS implemented in Python (not recursive CTE) for predictable node-count cap: scidex/atlas/relationship_graph.py with get_neighborhood() and find_shortest_path().
• Added relationship_view to ARTIFACT_TYPES in artifact_registry.py.
• Three API routes added to api.py:

- GET /relationship-graph/{artifact_id} — HTML page, injects ROOT_ARTIFACT_ID
- GET /api/relationship-graph/{artifact_id}.json — JSON BFS/path response
- POST /api/relationship-graph/save-view — register view as artifact

• site/relationship-graph.html: SVG D3 force-directed graph; link-type colored; node-shape by type (diamond=hypothesis, square=analysis/notebook/dataset, circle=default); click→artifact page; hover tooltip; depth/direction/link-type controls; shortest-path highlighted in gold.
• 9 pytest tests in tests/atlas/test_relationship_graph.py — all pass.

Human acceptance verified:

• depth=3 from top AD hypothesis (hypothesis-h-var-de1677a080) returns 200 nodes (truncated), 12 distinct link types — well above the ≥10 nodes / ≥3 types threshold.