ArcRAG Tests Whether AI Can Search Differently Built Knowledge Graphs

The question behind the graph

A question might require identifying a film’s cast, checking which cast members have children, and determining where those children studied. The ArcRAG paper uses this kind of example to explain why retrieving individually relevant passages differs from explicitly following relationships between facts.

Knowledge graphs represent entities and their connections. But those connections can be stored differently: one construction might use predefined relationship types, while another records freely worded descriptions. The authors ask whether a retrieval procedure can work across these differences without construction-specific prompts, relation names, or retuning.

In their November 4, 2026 preprint, Jyotin Goel and colleagues at Domyn introduce ArcRAG to investigate that question. Their central finding is bounded: one retrieval procedure worked across the two graph-construction families they tested, but accuracy still depended on construction.

Discovering the local map

ArcRAG first reads the graph’s storage schema: its node labels, relationship types, and property keys. This reveals how information is organized, but not necessarily which relationships are relevant to a particular question.

The system then combines dense and keyword-based retrieval to find starting points. Two exploration agents query the graph around those anchors. They report nearby relationship descriptions, paths, and possible naming variants. The paper calls this report a query-local interface: a description of relevant structure discovered for the current question, rather than a manually supplied vocabulary.

A planner selects candidate traversals through that structure. A generator turns them into read-only Cypher queries, a language for querying graph databases. When a query produces an unacceptable result, such as an error or an empty response, the generator can revise predicates and bindings while retaining the traversal’s shape.

An answer model composes a response from the retrieved evidence and judges whether it is sufficient. If not, a follow-up question can start another pass, with evidence accumulating across passes. These checks guide search; they do not establish that an answer is correct.

Two constructions, three benchmarks

The researchers built two graphs from each of three question-answering corpora: MuSiQue, HotpotQA, and TriviaQA. MuSiQue includes questions requiring up to four linked steps; HotpotQA is predominantly two-step; TriviaQA is predominantly single-step.

The schema-guided construction, based on FinReflectKG, used typed nodes, a closed relationship vocabulary, and pointers back to source passages. The open-information-extraction construction, based on HippoRAG’s methodology, used a single node label, freely worded predicates stored on edges, and no source-passage pointers.

Within each dataset, ArcRAG’s prompts, models, and hyperparameters stayed fixed across both constructions. The accompanying evidence store changed according to provenance: passage-backed graphs supplied source passages, while graphs without those pointers supplied retrieved graph rows. This distinction matters because the comparison changes both graph structure and the evidence available for answering.

What the reported scores show

On schema-guided graphs, ArcRAG achieved 44.0% exact match on MuSiQue and 74.5% on HotpotQA. On open-vocabulary graphs, those scores were 38.0% and 65.5%. Exact match measures whether a normalized answer matches a reference answer. On TriviaQA, token-level F1—which measures answer-word overlap—was 87.9% and 71.6%, respectively.

The authors interpret these results as evidence of portability, not unchanged performance. The system operated across both constructions, but the schema-guided graphs produced higher scores.

Component experiments on a stratified sample of 100 questions offer more direct evidence about the mechanism. Adding exploration to the generator improved exact match at every tested retry-and-pass setting: by 15–39 percentage points on schema-guided graphs and 12–32 points on open-vocabulary graphs.

Planning was less consistently helpful. At five retries and five passes on the schema-guided graph, exploration plus generation reached 88% exact match, versus 66% with planning added. These are small-sample component results, not the main benchmark scores. They also compare matching retry-and-pass budgets, not equal total inference cost: extra stages require extra model calls.

Where the evidence stops

The study tests only two language-model-extracted, entity-centered graph families. It does not establish transfer to curated ontologies or graphs whose nodes represent different kinds of units. It also does not isolate how much the accuracy gap comes from structure versus source provenance, or measure how much naming noise the interface can absorb.

Each configuration ran once. The answer model used temperature 1.0 and judged its own evidence sufficiency; run-to-run variation was not quantified. Pretrained knowledge may also contribute to scores on these public benchmarks.

Comparisons with previously published systems use different models and protocols, so they do not establish controlled superiority. Finally, the authors do not verify claim-level attribution. An executable evidence path makes retrieval inspectable, not automatically factual. Despite the paper’s motivation in regulated domains, its encyclopedic experiments do not establish reliability in finance, law, healthcare, or compliance.