Tutorial: Build a RAG Agent

This tutorial builds a RAG Agent from a set of product documents, focusing on verifying chunking, hybrid retrieval, reranking, citations, and failure diagnostics.

Final Result

The Agent answers questions from the documents and returns sources; the retrieval lab explains each result's scores across the vector, full-text, hybrid, and rerank stages.

Step 1: Create a Knowledge Base & Select Models

1. Create a knowledge base and select the current team.
2. Select the embedding model and optionally a rerank model.
3. Use a chunk size of 1000 and an overlap of 100.
4. Save.

Step 2: Import & Process Documents

1. Select Upload documents and upload the product document.
2. After the document enters Pending, select Preview chunks.
3. Check the chunk count, overlap length, and text cleaning results.
4. Select Start processing and wait until the status becomes Completed.

Step 3: Run the Retrieval Lab

1. Open Hit testing and enter a question with a clear answer in the documents.
2. Run vector, full-text, and hybrid retrieval separately.
3. Compare result overlap counts, per-stage latency, final scores, and diagnostics.
4. Save the best-performing local preset.
5. Before selecting Apply to production, confirm at least one hybrid weight is greater than 0.

Step 4: Create & Link the Agent

1. Create an Agent and select the chat model.
2. Link the knowledge base you just created and use hybrid.
3. Select Agentic RAG or Naive RAG as the RAG mode.
4. Prompt: “answer based only on sources; state uncertainty when no answer can be found”.
5. Preview and publish.

Acceptance & Troubleshooting

• Document questions cite the correct documents.
• Unanswerable questions do not produce unsupported claims.
• Retrieval lab diagnostics are empty or explainable.
• Rebuild the index after switching embedding models.
• If recall fails, check Qdrant, PostgreSQL/pg_search, model credentials, and document status.