Skip to content

Ontology-Anchored Path Resolution: Reducing Traversal Cost in AI Goal Systems

Status: early draft — outline settled, case study grounded, literature threads open


Thesis

AI systems currently resolve user goals through broad path traversal: vector similarity search, web retrieval, and long-context scan over candidate corpora. The cost of this traversal — measured in tokens, API calls, latency, and energy — scales with the size of the information space, not the complexity of the goal. When a goal can be mapped to a concept already represented in a public ontology or structured corpus network, the traversal collapses to a directed graph walk whose depth is bounded by the ontology's structure, not by the corpus size. Ontological anchoring of user intent converts O(n) scan into O(depth) lookup, and where full lookup is impossible, ontological context produces focus-bounded scan that is one to two orders of magnitude cheaper than blind scan. This paper formalizes the claim, analyzes the efficiency gains, and demonstrates the principle through the yidam architecture — a git-native knowledge graph system in which agents navigate structured corpus networks by explicit typed edges rather than by retrieval.


Paper sections

#TitleStatus
1The traversal problemopen
2Ontology and corpus networks as mapsopen
3Focused scan as the fallback regimeopen
4System architecture: yidam as case studydrafted
5Efficiency analysis and implicationsopen

Full section-by-section structure with key claims: outline.md


What this paper is not

  • Not a systems paper proposing a new retrieval algorithm
  • Not a benchmark paper — though it now carries one. yidam bench measures the §5 claim against a committed goal set, and yidam bench --scaling measures it over generated corpora. The generated corpora are circular by construction: the slope follows from the degree distribution chosen for them. Their parameters are derived from a real corpus, committed as configuration, and reported in every run, which makes the result arguable rather than neutral — see §5 of outline.md. The query surface the anchored arm needs is specified in RFC-0018, which takes the thesis below as its premise.
  • Not an argument against embedding-based retrieval — retrieval is the right fallback; the paper argues it should be a fallback, not the default

Key terms

Path traversal — the process by which an agent moves from a user's stated goal to a set of relevant knowledge artifacts. Encompasses retrieval, search, graph walk, and long-context reasoning over candidate material.

Ontological anchor — a mapping from a user's goal (expressed in natural language) to a node in a formal ontology or structured corpus. Once anchored, traversal proceeds along typed edges rather than by similarity.

Focused scan — a retrieval operation whose candidate set is constrained by ontological context to a subset of the full corpus; contrasted with blind scan (no ontological context) and full lookup (traversal replaces retrieval entirely).

Corpus network — a connected body of knowledge nodes with typed, directional edges, maintained with provenance. Distinguished from a document corpus by the presence of explicit structure.