Project Pitch:
I built a local-first, agentic RAG system that processes complex user queries by reasoning, researching, and validating across multiple specialized agents.
Instead of simple retrieval, the system splits questions into sub-tasks, routes each sub-question to a domain expert, verifies the outputs with a Supervisor Agent, and synthesizes a complete, high-quality final answer.
- Receives complex natural language questions
- Breaks down questions into sub-tasks using smart parsing
- Routes each sub-task to specialized domain agents
- Each agent retrieves and reasons about its specific topic
- A Supervisor Agent validates, merges, and finalizes the structured answer
- All data processing happens locally to protect privacy and improve speed 🧠 Supports multi-turn reasoning 📄 Handles structured and unstructured document input 🔒 Local-first processing, no cloud required
- n8n (workflow orchestration)
- Local LLM connections (Ollama / Custom LLM base URLs)
- Local vector search (PGVector, PostgreSQL)
- Metadata-based agent routing
- Multi-agent reasoning flow design
- RAG (Retrieval-Augmented Generation) strategies with local knowledge bases
- local_agentic_rag_system_customized.json — The customized n8n workflow
- local-agentic-rag-system-diagram.png — Visual system flow diagram
LARS demonstrates how AI systems can move beyond simple lookup responses — building local-first, secure, multi-agent reasoning systems that adapt to complex real-world user queries.
This is a blueprint for real-world, scalable AI deployment in enterprises where data privacy, modularity, and flexibility are critical.
If the online document retrieval API fails (due to a timeout or outage), the system activates a local fallback node.
This ensures the user still receives a valid response using cached knowledge.
Merged results are then passed forward to the Supervisor Agent for further processing.
- Add offline query fallback mode using cached vector store
- Fine-tune local LLMs for internal domain queries
- Create continuous learning pipeline to improve local knowledge base
- Log failed queries to detect scope gaps in local sources
Demo built for AI Agent Implementation Manager portfolio presentation.