Key Takeaway
Treat documents as small, typed knowledge graphs so AI agents query only relevant pieces—cutting token use by 60–95% while preserving or improving task accuracy.
ON THIS PAGE
Key Findings
Reformatting Markdown into a graph of semantic nodes with typed links lets agents fetch only the nodes relevant to a task and their declared dependencies. A tiny two-call protocol (index then resolve) lets agents traverse those graphs without loading entire files into the model context. Across a 240-document benchmark and 8 task types, this approach reduced token consumption dramatically and often improved accuracy by removing irrelevant content and adding structural cues. The format stays human-readable and can be produced from existing Markdown via a transpiler that copies content verbatim and uses models only for navigational metadata. Tree of Thoughts Pattern A2A Protocol Pattern
Data Highlights
1Overall token reduction of 60–95% across five document classes and eight task types.
2Operational runbooks saw up to 99.2% token reduction for single-question lookups (examples: 100k tokens → ~800 tokens).
3A 50-file skill library (~90,000 tokens) became navigable via a combined index of ~1,500 tokens (≈98.3% reduction in discovery cost).
What This Means
Engineers building autonomous agent pipelines: get much lower per-call costs and fewer context overflows by serving agent queries from a graph instead of injecting whole files. Technical leads and documentation owners: maintain a single human-readable source while enabling precise, role-scoped access for orchestrator and worker agents, improving reliability and auditability. Orchestrator-Worker Pattern
Not sure where to start?Get personalized recommendations
Ready to evaluate your AI agents?
Learn how ReputAgent helps teams build trustworthy AI through systematic evaluation.
Learn MoreYes, But...
Cross-file federation (resolving graph edges across repositories) remains an open problem and will be needed for large, distributed corpora. Evaluation used specific high-end models and a curated benchmark; results may vary with different models or real-world scale and workloads. The Markdown-to-graph transpiler keeps content verbatim but relies on model-generated navigation metadata, so authoring discipline and review remain important to avoid misrouting or stale scopes. Planning Pattern
Full Analysis
Agents today waste most of their input context by loading entire Markdown documents even when only a small section is relevant. Replacing flat files with ObjectGraph (.og) turns a document into a directed graph of typed semantic nodes and edges plus a tiny manifest. Agents use a two-primitive query protocol—an index pass to find candidate nodes and a resolve pass to fetch nodes plus declared dependencies and role-scoped content—so only needed text enters the model context. A2A Protocol Pattern Tree of Thoughts Pattern
The authors implemented a transpiler that converts ordinary Markdown into .og without rewriting content: deterministic parsers copy text verbatim while models generate only navigational metadata, limiting hallucination risk to pointers rather than facts. They evaluated on 240 documents (five classes) and eight task types using leading large models. Results showed 60–95% average token reductions and, in cases such as operational runbooks, reductions above 99% for single queries. Accuracy improved or held steady, attributed to reduced irrelevant noise and explicit semantic tags (for warnings, steps, assertions). The next steps are standardising cross-file edge resolution and integrating ObjectGraph with existing toolchains and access controls to support distributed knowledge graphs and governance. A2A Protocol Pattern
Explore evaluation patternsSee how to apply these findings
Credibility Assessment:
No recognizable affiliations or author reputation provided, arXiv preprint with zero citations; lacks identifiable institutional or established-author signals.