RAResearch AgentAI ENGINEERING LAB / 02
For recruiters — how to test and what to watch

What it is: a fully working agentic RAG system. You ask a question and the agent (an LLM with tools) plans its own searches, retrieves chunks from a vector database and writes an answer grounded only in them, with [n] citations. When the base lacks the information, it says so instead of making things up.

HOW TO TEST (30 S)

  1. Click an example question below or type your own.
  2. Send it and watch the “Agent trace” — sub-queries and retrieved chunks with similarity scores.
  3. Read the answer with [n] references and the “Sources” section.
  4. Optional: “Retrieval diagnostics” shows the raw vector search without the LLM.

WHAT IT DEMONSTRATES

  • Agent trace → question decomposition and multi-step retrieval (function calling)
  • Similarity scores → ANN vector search (cosine)
  • [n] citations + Sources → grounding, verifiability, anti-hallucination
  • Latency and tokens → cost and performance awareness

STACK & TECHNIQUES

Chunkingrecursive · ~1200 chars · overlap 180
EmbeddingsOpenAI text-embedding-3-small · 1536D
Vector databaseQdrant · Cosine · HNSW (ANN)
Agentgpt-4.1-mini · function calling · up to 5 turns · forced 1st retrieval
Groundinganswer only from context + [n] citations + refusal when missing
Retrievaltop-k 5–8 · cosine score
Observabilitytrace, scores, tokens, latency, server logs
Guardrailsrate limiting · token-gated ingest · graceful degradation
StackNext.js · Node runtime · Vercel · keys server-side only

Planned extensions: web search as another agent tool, reranking (cross-encoder), an evaluation set (recall@k, faithfulness).

ASK

Research agent

Pick a question and see the agent in action

Click one of the example questions above or type your own. You'll see, step by step, how the agent plans its searches, retrieves chunks from the vector database and builds an answer with citations to the sources.