Multi-Hop RAG at the Edge: Running 8 Agents Over a Local Knowledge Graph
The architecture behind Serpaix's multi-agent system: how Research, Graph, and Temporal agents collaborate over a shared memory store without hallucinating connections.
The Problem with Single-Pass RAG
Standard RAG does one retrieval, one generation. But complex questions — "What did I learn about transformers last month and how does it relate to my current project?" — require multiple reasoning hops across different parts of your knowledge base.
The Multi-Agent Architecture
Serpaix runs 8 specialized agents that collaborate over a shared memory store:
| Agent | Role |
|---|---|
| Research | Finds relevant notes across your knowledge base |
| Graph | Traverses the knowledge graph for 1-hop and 2-hop connections |
| Temporal | Orders results chronologically and identifies time-bound patterns |
| Contradiction | Detects conflicting information across sources |
| Summarizer | Condenses retrieved context into digestible form |
| Verifier | Cross-checks citations against source notes |
| Planner | Decomposes complex questions into sub-queries |
| Writer | Produces the final grounded answer |
Orchestration Flow
Preventing Hallucination
The key insight: by separating retrieval, verification, and generation into distinct agents, each with a specific narrow role, we dramatically reduce the risk of hallucination. The Writer can only use context provided by Research — it cannot invent facts. The Verifier independently checks every citation.
This architecture runs entirely at the edge - all agents execute within a 4-second budget, making it feasible for real-time chat. The same memory is exposed to external agents over the local MCP server, and the connections the Graph agent traverses come from the auto-built knowledge graph.