Why Vector RAG Fails Blast-Radius Questions (and the 4-Hop Cypher Query That Fixes It)

Description and use

Vector RAG fails on multi‑hop dependency questions; FalkorDB Cypher queries deliver deterministic, provenance‑backed answers for incident response.

Problem

Vector RAG cannot resolve multi‑hop dependency questions because it relies on semantic similarity rather than graph topology, leading to missed cascading failures during outages.

Outcome

The FalkorDB graph agent provided deterministic, provenance‑backed answers to blast‑radius questions, enabling rapid incident triage.

Workflow

• 1. Identify failing system (ledger‑service).
• 2. Run 4‑hop Cypher query to find downstream consumers.
• 3. Run expertise query to rank engineers by active project involvement.
• 4. Run decision provenance query to retrieve related ADRs and context.
• 5. Use results to inform incident response.

Evidence

Vector RAG operates by chunking text, generating high‑dimensional embeddings, and performing cosine similarity search.

Source

What it cannot do is compute transitive closure across an architecture diagram.

Source

The graph is load‑bearing. The LLM is only a router and explainer. Every answer must be a deterministic Cypher traversal with verifiable provenance.

Source

When ledger‑service fails, what breaks downstream? We execute a variable‑length incoming traversal (-[:DEPENDS_ON*1..5]-).

Source