"GraphRAG Iterative Retrieval Research 2026"

GraphRAG Iterative Retrieval Research (2026)

1. The Multi-Hop Retrieval Problem

Single-pass RAG frequently fails on complex queries where evidence for the answer is not directly in the query but is connected through intermediate entities (A → B → C).

2. The Retrieve-Reason-Retrieve Loop

Vox adopts an iterative loop for high-complexity queries:

  1. Initial Retrieval: Standard hybrid search over Tier 1/2.
  2. Partial Synthesis: Socrates (or Lane G) identifies missing constraints.
  3. Query Expansion: vox-search generates refined sub-queries based on partial evidence.
  4. Re-Retrieval: Fetches new evidence without duplicating existing fetches.
  5. Final Synthesis: Unified Socrates gate pass.

3. Key Heuristics

3.1 Stopping Conditions

  • evidence_quality ≥ 0.85.
  • Max hops reached (default: 3).
  • Zero unique URLs returned in the latest hop.

3.2 Constraint-Checked Retrieval (C2RAG)

Decomposes the query into atomic constraints. Before synthesis, the system verifies that each constraint has at least one supporting chunk in the corpus. Missing constraints trigger a targeted research hop.

4. Performance Impacts

Iterative loops increase total research latency by 2x-3x. This is gated by the Orient Phase; only tasks in the HighRisk or MultiHop complexity band trigger expansion.

5. References

  • HippoRAG: Knowledge Graphs for Collaborative Reasoning (2024)
  • GraphRAG-rs Technical Spec (2026)