Building AI agents that actually help a business is hard because context is scattered. Allevia, a project by Ankit Vohra, tackles this by giving each business its own FalkorDB graph, effectively creating a 'company brain' that 13 specialized AI agents must consult before acting. Built in just four days for the Graph Hacks hackathon, the system demonstrates that vector search alone is insufficient for answering complex operational questions like 'Why did online orders drop?'

The Graph as the Source of Truth

The core architecture isolates each business in its own graph, named allevia_, ensuring no cross-tenant data leakage. Instead of relying on a single vector database for retrieval, Allevia uses a hybrid approach: vector and full-text search identify starting points, while Cypher queries walk the actual relationships to find causal chains. For example, a drop in orders isn't just a number; it's a path from a payment error to a specific code release and the engineer responsible. This structure allows the agents to explain the 'why' rather than just reporting the 'what'.

Multi-Agent Coordination and Safety

The system employs 13 distinct agents, including a 'Chief' for routing, a 'Data Detective' for analysis, and a 'Fact Checker' for validation. These agents operate on the Claude Agent SDK and run in parallel, with each producing a 'Finding' node that includes citations. Crucially, the system enforces strict safety rails: every write operation passes through a gate that requires a source and a date, and every answer is checked against the tools' actual outputs. If a fact changes, the old edge is closed with a timestamp rather than deleted, preserving a temporal history of business decisions.

Performance and Token Efficiency

Practical benchmarks show significant efficiency gains. In a test using real data from the startup Aamdani, answering two complex questions required 75,043 tokens when using the graph tools, compared to 603,665 tokens when pasting all 114 business documents into the prompt. This represents an 8x reduction in token usage and a 43% decrease in response time. The graph queries themselves are fast, with shortest-path calculations taking between 5 and 218 milliseconds on a laptop running FalkorDB 6.0.2, proving that the database is not the bottleneck; the LLM inference is.

Key Takeaways

  • Hybrid retrieval (vector + graph) is essential for causal reasoning in business AI.
  • Temporal edges with valid and record timestamps prevent agents from being confidently wrong about outdated facts.
  • Graph-based context reduces LLM token costs by up to 8x compared to document dumping.
  • Strict write gates and citation verification are critical for autonomous agent reliability.

The Bottom Line

Vector search is a lookup tool, but graphs are reasoning tools; Allevia proves that without a graph to map causal relationships, AI agents remain reactive dashboards rather than proactive business partners.