Project Deep Dive
Notion-Style AI Knowledge Graph & RAG Engine
An end-to-end RAG pipeline that turns unstructured documents into structured, queryable knowledge graphs.
Overview
This Boston University project focused on converting unstructured documents into structured knowledge graphs using an end-to-end RAG pipeline. The system combined semantic querying, vector embeddings, and low-latency inference orchestration to make large document collections easier to explore and retrieve from.
Key Outcomes
Reduced retrieval time by about 75% across 15,000+ lines of content.
Built an end-to-end RAG pipeline for turning unstructured documents into structured knowledge graphs.
Orchestrated Dockerized multi-model inference services with gRPC for low-latency communication between the UI and AI layers.
Architecture Diagram
Documents are transformed into graph and vector representations, then retrieved through orchestrated inference services.
Architecture
Document ingestion transforms unstructured content into graph-friendly entities and relationships for downstream retrieval.
Semantic querying and vector embeddings run on PostgreSQL with pgvector to support fast similarity search across the knowledge base.
Llama 3 and LangChain orchestrate retrieval and answer generation on top of the structured graph plus vector store.
Dockerized multi-model inference services communicate with the UI and AI orchestration layer over gRPC for low-latency responses.
Stack
- Python
- LangChain
- Llama 3
- PostgreSQL
- pgvector
- Docker
- gRPC