Week 13 — Knowledge Graphs as Structured Agent Memory

Retrieval · Knowledge Graph · Agents

Advanced

V — Graph engineering

Overview

Switch from model internals to the information system around the model. Facts become explicit triples; retrieval becomes path search as well as semantic similarity. The eight-layer pipeline: ingestion → extraction → entity resolution → graph storage → retrieval → agent → verification → update.

What You Will Learn

• Represent evidence as subject–relation–object triples.
• Explain why entity resolution must precede scale.
• Build vector + entity + path + temporal retrieval.
• Answer a multi-hop question via a graph path.

Core Concepts

Triples

\text{Subject}\rightarrow\text{Relation}\rightarrow\text{Object}

Eight-layer pipeline

ingestion, extraction, entity resolution, graph storage, retrieval, agent, verification, update.

Hybrid retrieval

Vector lookup + entity lookup + path search + community retrieval + temporal filtering.

Prerequisites

Work these pages on this site before the lecture.

• BM25
• TF-IDF
• Precision / recall at K
• NDCG
• Graph representations
• Random walks & PageRank

Lecture notes

The eight layers are a pipeline, not a slogan: ingest → extract triples → resolve entities → store → retrieve (vector + entity + path + temporal) → agent → verify → update. Skipping resolution splits one real-world entity into two nodes and every multi-hop path dies.

A path beats a vector on “why / who / before” questions; a vector wins on fuzzy paraphrase. Your lab must show one question of each kind, including a case where the graph loses.

Required readings

• From Local to Global: GraphRAG
• Neo4j graph concepts · Cypher manual
• Random walks / PageRank

Related on this site

• Graph representations
• NDCG · BM25
• Week 14

Lab / Implementation

Convert a small document collection into entity/relation/evidence records and load them into Neo4j; implement vector + entity + path retrieval (Lab 1) and answer one multi-hop 'why?' via a path (Lab 2).

Not on this site (paper reading required)

• triples/extraction
• entity resolution
• the 8-layer pipeline
• Cypher/Neo4j
• grounding

Mastery Check

• DERIVE: the eight-layer pipeline; when a path beats a vector
• IMPLEMENT: docs→triples→resolution→Neo4j→hybrid retrieval
• BENCHMARK: multi-hop accuracy, path correctness, latency
• DEBUG: an entity-resolution miss splitting a node
• EXPLAIN: why resolution must precede scale

Part of CS/AI 684 — Efficient Long-Context Agent Systems.