Concepts¶
The ideas behind Odin: what it does, how the retrieval pipeline is assembled, and how each scoring signal contributes to the final ranking.
Odin's job is narrow and deliberate: given seed entities, return the most relevant scored paths through a knowledge graph. It does not answer natural-language questions or generate prose; that is the agent's job. Odin is the compass; the agent is the explorer.
-
Architecture
How PPR, beam search, NPLL, and aggregation compose into one pipeline.
-
Data Model
Entities, relationships, and communities: how Odin sees your graph.
-
Personalized PageRank
Finding the structurally important nodes relative to your seeds.
-
Beam Search
Bounded, best-first multi-hop path exploration.
-
NPLL Edge Scoring
Learned edge plausibility that filters invalid paths.
-
Motifs & Aggregation
Turning paths into recurring patterns and summaries.
-
Triage & Insight Scoring
Collapsing many signals into one 0-100 prioritization number.
-
Caching
Why graph access is LRU-cached and what it costs.
-
Adapters
Connect Odin to ArangoDB, JanusGraph, or your own backend.
The three signals¶
Every path Odin returns is scored by combining three complementary signals:
| Signal | Question it answers | Component |
|---|---|---|
| Structural | Is this node important in the graph? | Personalized PageRank |
| Reachable | Can we get there efficiently? | Beam Search |
| Semantic | Is this edge plausible? | NPLL |
No single signal is sufficient. PPR alone finds important nodes but follows nonsensical edges. Beam search alone explodes without a scoring signal to prune. NPLL alone validates edges but has no notion of importance. Together they let an agent focus on paths that are important, reachable, and plausible.