GraphContainer: One Platform to Compare and Debug Graph RAG
Ask any team running Graph RAG in production which subgraph traversal actually fired on last night's failing multi-hop question. Silence.
GraphContainer: One Platform to Compare and Debug Graph RAG
Ask any team running Graph RAG in production which subgraph traversal actually fired on last night’s failing multi-hop question. Silence. The retrieval trace is buried inside a bespoke pipeline, the graph format is proprietary, and swapping one method for another means a two-week rewrite.
That’s the gap GraphContainer, accepted as a demonstration at VLDB 2026, is built to close.
What the platform actually does
GraphContainer introduces a Unified Graph Representation layer that normalizes heterogeneous graph formats — property graphs, RDF, hypergraphs, document-derived KGs — into one internal schema. On top sits a Graph Recorder that captures every retrieval step and replays it in a web UI, so an engineer can inspect which nodes were touched, which edges were pruned, and which hop returned no candidate.
For teams building retrieval-augmented systems, this changes the debugging economics. You compare two methods at the level of the traversal itself, not just the final answer.
POST /gc/session/replay
{
"query": "Which board members of X also sit on Y's audit committee?",
"method": "gnn-retriever-v2",
"record": true
}
→ 200 OK
trace_id: gc-2026-07-24-a91f
hops: 3
nodes_touched: 47
edges_pruned: 128
final_context_tokens: 812That kind of trace turns “the answer was wrong” into “hop 2 selected the wrong entity because the disambiguation score fell below threshold.”
Why standardization matters here
Graph RAG research has fragmented fast. Methods now range from community-based summarization (see Microsoft’s GraphRAG project) to path-ranking retrievers built on Neo4j and other property-graph backends. Each ships its own loader, its own evaluation harness, and its own definition of “context.”
Without a common substrate, benchmarks like HotpotQA produce numbers that aren’t strictly comparable across pipelines. GraphContainer’s contribution is not a new retriever — it’s the shared bench that lets you run three of them against the same graph and see which one hallucinates a bridging entity.
Where this sits in a production stack
At CAI Technology we treat Graph RAG traceability as a control problem, not a UX feature. A retrieval pipeline you cannot replay deterministically is one you cannot audit under Directive (EU) 2022/2555 (NIS2) or defend under a GDPR Art. 22 challenge. GraphContainer’s Recorder is the kind of primitive that makes those audits mechanical. It also aligns with ENISA guidance on AI system observability: if you cannot show which data influenced an output, you cannot defend the output.
Our own bias is to run the Recorder in shadow mode next to production, sampling a thin slice of live queries into an offline comparison harness. When a graph update silently changes retrieval behavior, the trace diff surfaces it before the accuracy metric moves.
If you’re wiring Graph RAG into a regulated workflow, our detection and control patterns show where the Recorder fits alongside the rest of the stack.