RAG Architecture Example

Spring Boot application that demonstrates different Retrieval-Augmented Generation (RAG) patterns using Spring AI, Ollama, and PostgreSQL pgvector.

Tech Stack

• Java 25 with virtual threads
• Spring Boot 4.0.3 + Spring AI 2.0.0-M2
• Ollama — granite4:3b (chat) and granite-embedding:278m (embeddings)
• PostgreSQL + pgvector — vector store with HNSW indexing
• Docling Serve — PDF document parsing and chunking
• Arconia 0.22.0 — dev services and Docling integration
• OpenTelemetry — observability via Grafana LGTM stack

RAG Pipelines

The project showcases four RAG strategies through different endpoints:

1. Simple RAG (GET /chat?question={query})

Basic retrieval-augmented generation. Searches the vector store for relevant documents and augments the LLM response.

2. Multi-Query RAG (POST /rag/query/multi-query)

Expands the user query into 3 variations using MultiQueryExpander, executes them in parallel, and combines the retrieved documents for a more comprehensive answer.

3. Optimized RAG (POST /rag/query/optimization)

Translates the query to English (supports multilingual input), then applies multi-query expansion for improved retrieval quality.

4. Query Routing RAG (POST /rag/query/routing)

Agentic approach where the LLM selects the appropriate retrieval tool based on query semantics:

• Iorek & Pingu Retriever — searches North Pole story documents
• Lucio & Balosso Retriever — searches Italian Alps story documents
• Web Search Retriever — searches the web via Tavily API

Document Ingestion

Two ingestion pipelines run automatically on startup:

Pipeline	Source	Strategy
IngestionPipelineDocling	Designing Machine Learning Systems (PDF, ~390 pages)	Splits into 5-page batches, processes via Docling Serve
IngestionPipelineBranching	story1.md, story2.md	Markdown reader with location metadata tagging

Prerequisites

• Docker Desktop (for dev services: Docling, PostgreSQL, Ollama, Grafana LGTM)
• Java 25+
• Tavily API key (optional, for web search routing) — set TAVILY_SEARCH_API_KEY env var

Running

./gradlew bootRun

Dev services (Docling Serve, PostgreSQL + pgvector, Ollama, Grafana LGTM) start automatically via Arconia/Testcontainers.

Project Structure

src/main/java/com/marcos/ragarchitectureexample/
├── RagArchitectureExampleApplication.java
├── ragDocling/
│   ├── ChatController.java              # Simple RAG endpoint
│   └── IngestionPipelineDocling.java     # PDF ingestion via Docling
├── ragBranchingOllama/
│   ├── RagControllerMultiQuery.java      # Multi-query RAG endpoint
│   ├── RagControllerOptimization.java    # Translation + expansion RAG endpoint
│   └── IngestionPipelineBranching.java   # Markdown ingestion with metadata
└── ragQueryRouting/
    ├── RagControllerQueryRouting.java     # Tool-based routing RAG endpoint
    └── SearchEngineDocumentRetriever.java # Tavily web search integration

Configuration

Key settings in application.yaml:

Property	Value	Description
spring.ai.ollama.chat.options.model	granite4:3b	Chat model
spring.ai.ollama.embedding.options.model	granite-embedding:278m	Embedding model
spring.ai.vectorstore.pgvector.dimensions	768	Embedding dimensions
spring.ai.vectorstore.pgvector.index-type	HNSW	Vector index type
arconia.docling.read-timeout	300s	Docling API read timeout