Skip to content

confidence.md

Latest commit

 

History

History
109 lines (77 loc) · 4.11 KB

confidence.md

File metadata and controls

109 lines (77 loc) · 4.11 KB

Edge Confidence Labels

Every relationship in the codegraph property graph carries two properties:

  • confidence — categorical label: EXTRACTED, INFERRED, or AMBIGUOUS
  • confidence_score — numeric score in the range [0.0, 1.0]

These fields let consumers filter edges by reliability. A high-confidence subgraph (score >= 0.9) is useful for strict architecture checks, while the full graph including lower-confidence edges is better for exploratory analysis.

Taxonomy

Label Meaning Typical score
EXTRACTED Derived directly from the AST with no heuristics. The relationship is a syntactic fact. 1.0
INFERRED Resolved via a heuristic (name matching, barrel re-exports, MRO guessing). Likely correct but not guaranteed. 0.5 -- 0.9
AMBIGUOUS Multiple equally valid targets exist; the edge picks one but the choice is uncertain. < 0.5

AMBIGUOUS is reserved for future use. Today all edges are either EXTRACTED or INFERRED.

Edge Classification

Structural edges (always EXTRACTED / 1.0)

These come straight from the AST with no resolution step:

DEFINES_CLASS, DEFINES_FUNC, HAS_METHOD, DEFINES_IFACE, DEFINES_ATOM, HAS_COLUMN, EXPOSES, RESOLVES, BELONGS_TO, MEMBER_OF, OWNED_BY, LAST_MODIFIED_BY, CONTRIBUTED_BY, READS_ATOM, WRITES_ATOM, READS_ENV, HANDLES_EVENT, EMITS_EVENT, DECORATED_BY

Import edges (confidence varies by resolution strategy)

Strategy Label Score When
direct EXTRACTED 1.0 Resolved to an exact .py / .ts file
relative EXTRACTED 1.0 Relative import (e.g. from .b import x) resolved to a file
alias INFERRED 0.9 Resolved via a tsconfig.json path alias
workspace INFERRED 0.85 Resolved via workspace/monorepo package boundary
barrel INFERRED 0.8 Resolved through an __init__.py or index.ts barrel

External (unresolved) imports default to EXTRACTED / 1.0 -- the import statement itself is a fact.

CALLS edges (confidence varies by receiver kind)

Receiver Label Score Example
self / this EXTRACTED 1.0 self.foo(), this.bar()
super() INFERRED 0.7 super().run() (MRO guess)
this.field INFERRED 0.6 DI-injected field call
name (class target) INFERRED 0.5 Bare helper() resolved to a class method
name (function) INFERRED 0.5 Bare helper() resolved to a module function

Other cross-file edges

Edge kind Label Score Rationale
EXTENDS, IMPLEMENTS EXTRACTED 1.0 AST heritage clause
INJECTS, PROVIDES, EXPORTS_PROVIDER EXTRACTED 1.0 NestJS decorator metadata
CALLS_ENDPOINT INFERRED 0.7 URL pattern matching
RENDERS INFERRED 0.8 JSX component name resolution
USES_HOOK EXTRACTED 0.9 Hook name from AST
TESTS INFERRED 0.5 File-name pairing heuristic
TESTS_CLASS INFERRED 0.6 Test-class name matching

Querying Confidence in Cypher

High-confidence CALLS only:

MATCH (a)-[r:CALLS]->(b)
WHERE r.confidence_score >= 0.9
RETURN a.name, b.name, r.confidence, r.confidence_score

Distribution of confidence labels:

MATCH ()-[r]->()
WHERE r.confidence IS NOT NULL
RETURN type(r) AS edge_type, r.confidence, count(*) AS cnt
ORDER BY edge_type, r.confidence

Find all low-confidence edges:

MATCH (a)-[r]->(b)
WHERE r.confidence_score < 0.7
RETURN type(r) AS kind, a.name, b.name, r.confidence_score
ORDER BY r.confidence_score
LIMIT 25

Filtering in Arch-Check Policies

When writing or customising arch-check policies, you can exclude low-confidence edges to reduce false positives. For example, to check import cycles only among high-confidence imports:

MATCH path = (a:File)-[:IMPORTS*2..6]->(a)
WHERE ALL(r IN relationships(path) WHERE r.confidence_score >= 0.8)
RETURN [n IN nodes(path) | n.path] AS cycle

This filters out barrel-resolved imports (score 0.8) and below, keeping only direct and relative imports in the cycle analysis.