Scoring Edges

Retrieval walks whole paths, but sometimes you only need a verdict on one relationship. score_edge() gives you exactly that: the NPLL plausibility signal Odin uses internally, exposed so you can drop it straight into your own agent logic.

Calling it

You pass a source, a relation, and a destination, and get back a probability:

score = engine.score_edge(src, rel, dst)

Argument	Meaning
src	Source entity ID
rel	Relation type
dst	Destination entity ID

The result is a float between 0.0 (impossible, or contradicts the patterns Odin learned from your graph) and 1.0 (highly plausible). The difference is easiest to see with two edges from the same entity:

engine.score_edge("entity/patient_001", "treated_by", "entity/doctor_smith")
# 0.91   plausible: patients are treated by doctors
engine.score_edge("entity/patient_001", "diagnosed_by", "entity/aspirin")
# 0.03   implausible: a patient is not diagnosed_by a medication

Gating a decision

The typical use is a plausibility gate in front of an action. A high score is worth acting on, a very low score is worth discarding, and the uncertain middle is exactly where a human belongs:

def maybe_investigate(agent, src, rel, dst):
    score = engine.score_edge(src, rel, dst)
    if score > 0.7:
        agent.investigate_further(src, rel, dst)
    elif score < 0.2:
        agent.discard(src, rel, dst, reason="implausible edge")
    else:
        agent.request_human_review(src, rel, dst, confidence=score)

The same call is also the fastest way to validate a relationship an agent proposes but has not confirmed in the graph, such as checking an LLM's guess before you write it back:

proposed = ("entity/fund_A", "managed_by", "entity/sanctioned_entity")
if engine.score_edge(*proposed) > 0.5:
    compliance_agent.flag_for_review(proposed)

Cost and availability

Edge scores are cached (see Caching), so scoring the same triple repeatedly within a session is effectively free after the first lookup, at roughly 5 ms per edge. One thing to check before you lean on fine-grained plausibility: if the engine is in constant-confidence mode, score_edge() returns the constant fallback rather than a learned score.

if engine.has_npll:
    score = engine.score_edge(src, rel, dst)
else:
    ...   # no trained model yet; see the Model Lifecycle guide

If you find yourself in that mode, Model Lifecycle explains how to get a model trained.

Next

The concept behind the score is NPLL Edge Scoring, and AI Agent Integration shows score_edge() inside complete agent loops.