Title: Grokers: Bottom-Up Inductive Comprehension and Write-Time Intelligence over Typed Knowledge Graphs

Abstract

This paper introduces Grokers, a novel architecture designed to construct persistent, structured understanding of typed knowledge graphs via the inductive traversal of dependency subgraphs from the bottom up. In contrast to Retrieval-Augmented Generation (RAG), which incurs the full cost of comprehension for every individual query, Grokers shifts this intelligence to the write phase. Autonomous Groker agents examine nodes within a typed stream graph, utilizing governed language model (LM) calls to extract structured attributes. This understanding is then inductively composed upward through dependency relations, resulting in enriched typed attributes that satisfy all subsequent queries without any additional LM expenditure.

We formally demonstrate three key properties: 1. The Byte-Identity Theorem: This establishes that context blocks, constructed from a transactionally maintained denormalization index, remain byte-identical across LM turns as long as no semantic changes occur. This consistency allows for KV-cache hit rates nearing 100%. 2. The Accumulation Monotonicity Theorem: This proves that, under a governed wisdom library growth protocol, the proportion of interactions resolved without invoking LM calls is non-decreasing relative to the number of completed interactions. 3. The Dual-Traversal Ordering Theorem: This confirms that top-down generation and bottom-up comprehension represent the unique correct traversal orderings for their respective tasks within a dependency DAG. Furthermore, it shows that combining these two methods forms a complete cycle of generation and comprehension.

Additionally, we offer a deterministic alternative to embedding-based semantic search, featuring a synonym caching protocol. In domains with finite vocabularies, the LM fallback rate for this protocol converges to zero. A reference implementation is available within the open-source Qbix / Safebox / Safebots stack.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC