The Knowledge Layer

Evaluation: how do you know your agent setup is working?

Every technique in this series makes your knowledge layer better, or claims to.

Oct 6 • Ratko Nikolić

September 2026

Knowledge editing and why it isn't the answer for enterprise updates

There’s a research line on editing facts directly into a model’s weights: ROME, MEMIT, GRACE, WISE.

Sep 22 • Ratko Nikolić

Context engineering: the discipline that replaces prompt engineering

The model doesn’t see your corpus, your memory, or your tools: it sees a context window you assembled.

Sep 8 • Ratko Nikolić

August 2026

Vector vs graph vs episodic: a tour of agent memory systems

RAG retrieves from a fixed corpus. Memory is the part that remembers what happened, and the vendors building it have each bet on a different…

Aug 25 • Ratko Nikolić

Serbian Legal AI: a RAG architecture comparison

Five knowledge-layer architectures, one Serbian-law corpus, 300 graded answers.

Aug 11 • Ratko Nikolić

July 2026

Agentic RAG: when retrieval becomes a decision, not a pipeline

Naive and advanced RAG retrieve the same way every time. Agentic RAG lets the model decide what to look up, whether the results are good enough, and…

Jul 28 • Ratko Nikolić

GraphRAG and friends: when entities and relationships beat similarity

Some questions aren't about what a document says — they're about how things are connected. Vector search can't answer those. A typed knowledge graph…

Jul 14 • Ratko Nikolić

June 2026

Advanced RAG: what you actually run in production

Four moves turn the naive baseline into something that survives real traffic — hybrid search, reranking, query rewriting, and metadata filtering. Each…

Jun 30 • Ratko Nikolić

Naive RAG: the baseline you'll always benchmark against

The simplest retrieval pipeline there is — chunk, embed, match, stuff into the prompt. It's also the thing every fancier technique has to prove it…

Jun 16 • Ratko Nikolić

Your agent's problem isn't the model. It's the knowledge layer.

Most agent failures aren't model failures or prompt failures — they're failures in the layer underneath. This series is about that layer.

Jun 2 • Ratko Nikolić