spring-ai-playground

description: Tutorial 3 - upload a document, watch the ETL pipeline (extract → chunk → embed → store), and verify retrieval with a similarity search.

Tutorial 3 - Index a Document for RAG

Time 6 min · Difficulty ★☆☆ · Surfaces Vector Database

!!! abstract “Goal” Upload a document, watch it pass through the ETL pipeline (extract → chunk → embed → store), and verify retrieval quality with a similarity search before relying on it in chat.

Steps

  1. Open Vector Database. The header shows the active store and embedding model: SimpleVectorStore - Ollama: qwen3-embedding:0.6b.

Vector Database with the search controls and document sidebar ① indexed-doc sidebar (single source of truth), ② similarity-search input (hit Enter to query), ③ Spring AI metadata filter expression - same syntax you’d use in code.

  1. Click the document-add icon next to the sidebar to open the New Document & ETL Pipeline dialog. Drop in a PDF, DOCX, or PPTX - up to 20 MB.
  2. Tune the splitter only if the defaults don’t match your content shape. Chunk Size and Min Chunk Size Chars are the two that move retrieval quality the most.

New document dialog showing the splitter settings ① upload the file (drag-drop also works), ② token-splitter settings - Chunk Size and Min Chunk Size Chars move retrieval quality the most, ③ Chunk Document runs extraction + splitting and shows the chunks before embedding, so you can adjust the splitter without re-uploading.

  1. After embedding, run a similarity search to confirm retrieval works. Use a phrase that should be in the document.

Similarity search results with score, retrieved text, and metadata ① cosine similarity score (0.0-1.0), ② the retrieved chunk text, ③ metadata used by Spring AI filter expressions (source, chunk_index, custom fields).

!!! tip “Why this matters” Bad RAG starts here, not in chat. If the chunk you expect to be retrieved doesn’t show up here at a reasonable similarity score (≥ 0.6 for most cases), the chat answer will be ungrounded - no amount of prompting fixes that.

!!! warning “Don’t change the embedding model after indexing” The vector store stores raw vectors. Switching from qwen3-embedding:0.6b to a different model leaves the old vectors in place but indexed in a different space. Re-import or rebuild before trusting retrieval again.

This site is open source. Improve this page.