lee/graphify
1
0
Fork 0
mirror of https://github.com/safishamsi/graphify.git synced 2026-09-23 05:55:54 +00:00
Code Issues Packages Projects Releases Wiki Activity

Suppress cross-language INFERRED calls/uses edges in surprising connections

Browse Source 
In Python+TypeScript monorepos, the call and import resolvers match by
label across language boundaries (AuthError -> Member), producing false
positives that dominate 'Surprising Connections' due to cross-dir (+2)
and cross-community (+1) bonuses.

Expand the existing calls guard to also cover uses edges, and zero the
cross-dir and cross-community bonuses for these pairs — not just conf_bonus.
Leaves semantically_similar_to, EXTRACTED, and AMBIGUOUS edges unaffected.

5 new tests: calls suppressed, uses suppressed, semantically_similar_to
preserved, same-language INFERRED preserved, cross-language EXTRACTED preserved.
This commit is contained in:
Safi
2026-05-16 00:07:18 +01:00
parent fa70449d80
commit d14e8a72d1
2 changed files with 108 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files
graphify/analyze.py
+16 -5
View File
@@ -175,9 +175,20 @@ def _surprise_score(
relation = data.get("relation", "")
conf_bonus = {"AMBIGUOUS": 3, "INFERRED": 2, "EXTRACTED": 1}.get(conf, 1)
# Cross-language INFERRED calls are likely resolver pollution, not real surprises
if conf == "INFERRED" and relation == "calls" and _cross_language(u_source, v_source):
conf_bonus = 0 # downgrade: don't promote likely false positives
# Cross-language INFERRED calls/uses edges are resolver pollution in monorepos:
# the call and import resolvers match by label across language boundaries, so
# a Python `AuthError` resolves to a TypeScript `Member` purely by name.
# Zero all structural bonuses for these — they would otherwise score 4-5 from
# cross-dir + cross-community and dominate "Surprising Connections".
# Excludes `semantically_similar_to` (LLM-emitted, explicitly cross-language
# insight) and all AMBIGUOUS/EXTRACTED edges (not from the resolver path).
_suppress_structural = (
conf == "INFERRED"
and relation in ("calls", "uses")
and _cross_language(u_source, v_source)
)
if _suppress_structural:
conf_bonus = 0
score += conf_bonus
if conf in ("AMBIGUOUS", "INFERRED"):
@@ -191,14 +202,14 @@ def _surprise_score(
reasons.append(f"crosses file types ({cat_u} ↔ {cat_v})")
# 3. Cross-repo bonus - different top-level directory
if _top_level_dir(u_source) != _top_level_dir(v_source):
if _top_level_dir(u_source) != _top_level_dir(v_source) and not _suppress_structural:
score += 2
reasons.append("connects across different repos/directories")
# 4. Cross-community bonus - Leiden says these are structurally distant
cid_u = node_community.get(u)
cid_v = node_community.get(v)
if cid_u is not None and cid_v is not None and cid_u != cid_v:
if cid_u is not None and cid_v is not None and cid_u != cid_v and not _suppress_structural:
score += 1
reasons.append("bridges separate communities")
tests/test_analyze.py
+92
View File
@@ -123,6 +123,98 @@ def test_surprising_connections_cross_type_scores_higher():
assert any("code" in r and "paper" in r for r in reasons_cross)
def _make_cross_lang_graph():
"""Helper: Python node in backend/, TypeScript node in frontend/, different communities."""
G = nx.Graph()
G.add_node("py_auth", label="AuthError", source_file="backend/auth.py", file_type="code")
G.add_node("ts_member", label="Member", source_file="frontend/types.ts", file_type="code")
G.add_node("py_a", label="ServiceA", source_file="backend/service.py", file_type="code")
G.add_node("py_b", label="ServiceB", source_file="backend/utils.py", file_type="code")
return G
def test_cross_language_inferred_calls_suppressed():
"""Cross-language INFERRED calls edge should score lower than same-language EXTRACTED."""
G = _make_cross_lang_graph()
G.add_edge("py_auth", "ts_member", relation="calls", confidence="INFERRED",
weight=0.8, source_file="backend/auth.py")
G.add_edge("py_a", "py_b", relation="calls", confidence="EXTRACTED",
weight=1.0, source_file="backend/service.py")
nc = {"py_auth": 0, "ts_member": 1, "py_a": 0, "py_b": 0}
score_cross, _ = _surprise_score(G, "py_auth", "ts_member",
G.edges["py_auth", "ts_member"], nc,
"backend/auth.py", "frontend/types.ts")
score_same, _ = _surprise_score(G, "py_a", "py_b",
G.edges["py_a", "py_b"], nc,
"backend/service.py", "backend/utils.py")
assert score_cross <= score_same
def test_cross_language_inferred_uses_suppressed():
"""Cross-language INFERRED uses edge (the exact rsl-siege-manager false positive) should be suppressed."""
G = _make_cross_lang_graph()
G.add_edge("py_auth", "ts_member", relation="uses", confidence="INFERRED",
weight=0.8, source_file="backend/auth.py")
G.add_edge("py_a", "py_b", relation="calls", confidence="EXTRACTED",
weight=1.0, source_file="backend/service.py")
nc = {"py_auth": 0, "ts_member": 1, "py_a": 0, "py_b": 0}
score_cross, _ = _surprise_score(G, "py_auth", "ts_member",
G.edges["py_auth", "ts_member"], nc,
"backend/auth.py", "frontend/types.ts")
score_same, _ = _surprise_score(G, "py_a", "py_b",
G.edges["py_a", "py_b"], nc,
"backend/service.py", "backend/utils.py")
assert score_cross <= score_same
def test_cross_language_semantically_similar_not_suppressed():
"""`semantically_similar_to` across languages is a genuine insight — must not be suppressed."""
G = _make_cross_lang_graph()
G.add_edge("py_auth", "ts_member", relation="semantically_similar_to",
confidence="INFERRED", weight=0.85, source_file="backend/auth.py")
G.add_edge("py_a", "py_b", relation="calls", confidence="EXTRACTED",
weight=1.0, source_file="backend/service.py")
nc = {"py_auth": 0, "ts_member": 1, "py_a": 0, "py_b": 0}
score_sem, _ = _surprise_score(G, "py_auth", "ts_member",
G.edges["py_auth", "ts_member"], nc,
"backend/auth.py", "frontend/types.ts")
score_same, _ = _surprise_score(G, "py_a", "py_b",
G.edges["py_a", "py_b"], nc,
"backend/service.py", "backend/utils.py")
assert score_sem > score_same
def test_same_language_inferred_calls_not_suppressed():
"""INFERRED calls within the same language family must not be affected."""
G = nx.Graph()
G.add_node("py_a", label="ModuleA", source_file="src/a.py", file_type="code")
G.add_node("py_b", label="ModuleB", source_file="src/b.py", file_type="code")
G.add_node("py_c", label="ModuleC", source_file="src/c.py", file_type="code")
G.add_node("py_d", label="ModuleD", source_file="src/d.py", file_type="code")
G.add_edge("py_a", "py_b", relation="calls", confidence="INFERRED",
weight=0.8, source_file="src/a.py")
G.add_edge("py_c", "py_d", relation="calls", confidence="EXTRACTED",
weight=1.0, source_file="src/c.py")
nc = {"py_a": 0, "py_b": 1, "py_c": 0, "py_d": 1}
score_inf, _ = _surprise_score(G, "py_a", "py_b", G.edges["py_a", "py_b"], nc,
"src/a.py", "src/b.py")
score_ext, _ = _surprise_score(G, "py_c", "py_d", G.edges["py_c", "py_d"], nc,
"src/c.py", "src/d.py")
assert score_inf > score_ext
def test_cross_language_extracted_calls_not_suppressed():
"""EXTRACTED cross-language edges are real structural facts — must not be penalised."""
G = _make_cross_lang_graph()
G.add_edge("py_auth", "ts_member", relation="calls", confidence="EXTRACTED",
weight=1.0, source_file="backend/auth.py")
nc = {"py_auth": 0, "ts_member": 1}
score, _ = _surprise_score(G, "py_auth", "ts_member",
G.edges["py_auth", "ts_member"], nc,
"backend/auth.py", "frontend/types.ts")
assert score >= 1
def test_surprising_connections_have_why_field():
G = make_graph()
communities = cluster(G)