Introduction

Retrieval-Augmented Generation (RAG) has evolved significantly. While early systems relied on simple semantic search, modern architectures incorporate agentic behaviors—allowing models to reason about what to search for, verify the results, and iterate if necessary.

The Shift to Agents

Traditional RAG follows a linear path:

1. Retrieve documents
2. Stuff context
3. Generate answer

Agentic RAG introduces a loop:

1. Plan: Analyze the query.
2. Retrieve: Fetch data.
3. Critique: Is the data sufficient?
4. Refine: Search again if needed.
5. generate: Final answer.

// Example of a simple agent loop concept
async function agenticRetrieve(query: string) {
  const plan = await planner.generatePlan(query);
  let context = [];
  
  for (const step of plan) {
    const results = await vectorDb.search(step.query);
    if (evaluator.isRelevant(results)) {
      context.push(results);
    } else {
      // Re-write query strategy
    }
  }
  
  return generator.answer(query, context);
}

Conclusion

By giving LLMs tools and control flow, we move from static knowledge bases to dynamic reasoning engines.