Vector Databases vs Knowledge Graphs

Sample article — starter content for NeuralSys.

Introduction

"Should I use a vector DB or a knowledge graph?" is the wrong question. They answer different questions: what is similar? versus what is true and connected?

The Core Tradeoff

When Vectors Alone Fail

• Multi-hop questions ("services owned by teams that depend on X")
• Freshness and permissions ("only docs I may see, updated this quarter")
• Contradictory sources with no provenance

Embeddings will happily return a confident, similar, wrong answer.

The Pragmatic Hybrid

You don't need a full ontology on day one. Start here:

Documents → chunks → embeddings (vector DB)
Documents → entities + relations (graph, even a small one)
Query → retrieve from both → fuse → generate with citations

def hybrid_answer(query):
    vec_hits = vector_db.search(query, k=8)
    graph_hits = knowledge_graph.neighbors(query_entities(query), depth=2)
    evidence = fuse(vec_hits, graph_hits)
    return llm.generate(query, evidence, require_citations=True)

Key Takeaways

• Use vectors for similarity, graphs for structure — fuse at retrieval time.
• Start with a small, high-value graph (ownership, dependencies, lineage), not the universe.
• Provenance metadata is what makes either store production-worthy.