GraphMind — Amresh Chaurasiya

Overview

GraphMind bridges the gap between raw text and structured intelligence. Rather than relying purely on vector similarity, it extracts entities and relationships to construct a queryable knowledge graph. By combining vector search with graph traversal (Hybrid Retrieval), it provides evidence-backed reasoning that is highly resistant to hallucinations.

Architecture

01 — Documents

Raw unstructured input ingestion

02 — Parsing & Chunking

Semantic boundaries detection

03 — Extraction

LLM-driven entity and relationship mapping

04 — Knowledge Graph

Neo4j relational structure

05 — Vector Search

pgvector semantic embeddings

06 — Hybrid Retrieval

Graph + Vector scoring

07 — LLM Reasoning

Synthesizing evidence-backed answers

Why I built this

Standard RAG (Retrieval-Augmented Generation) fails on complex reasoning tasks because it lacks understanding of global relationships between concepts. I conceptualized GraphMind to solve the 'multi-hop' reasoning problem by explicitly defining relationships in a graph database.

Engineering Challenges

Accurately extracting strict semantic triplets (Subject-Predicate-Object) from messy text
Balancing latency between Neo4j graph traversal and PostgreSQL vector search
Evaluating the accuracy of AI-generated answers against the original source documents

What I learned

Hybrid retrieval (Graph + Vector) dramatically outperforms pure vector search when answering questions that require synthesizing information across multiple documents.

Stack

Next.jsFastAPIPython

PostgreSQLpgvectorNeo4j

Redis

Docker