Hybrid RAG approach combining AsciiDoc metadata with Docling HTML conversion.

This project uses Quarkus Docling to build a RAG (Retrieval-Augmented Generation) system for Quarkus documentation, producing a pgvector Docker image compatible with the chappie-server implementation.

This implementation combines the best of both approaches:

1. Git Clone - Clones Quarkus repository at specific version tag
2. Metadata Extraction - Extracts topics, categories, extensions, summary from AsciiDoc headers
3. HTML Fetching - Fetches versioned HTML from quarkus.io (e.g., /version/3.15/guides/)
4. Docling Conversion - Layout-aware HTML → Markdown conversion
5. Metadata Enrichment - Combines Docling content with AsciiDoc metadata
6. CDI Integration - Uses Quarkus CDI with fixed port (5001) for Docling service

• ✅ 100% RAG golden set accuracy (36/36 test cases)
• ✅ Rich metadata for improved search relevance
• ✅ Version consistency (metadata and content from same Quarkus version)
• ✅ Layout-aware conversion (preserves document structure)
• ✅ No GitHub rate limits (fetches from quarkus.io)

Git Clone → AsciiDoc Metadata + Versioned HTML → Docling → Markdown + Metadata →
Chunking/Embedding → PGVector → Docker Image

See APPROACH.md for detailed architecture and implementation notes.

• Java 21+
• Docker
• Maven 3.9+

./mvnw quarkus:dev

The Docling UI will be available at http://localhost:8080/q/dev/ for testing document conversions.

First, compile the application:

mvn clean package -DskipTests

Then run the bake-image command:

java -jar target/quarkus-app/quarkus-run.jar bake-image \
  --quarkus-version=3.15.0 \
  --chunk-size=1000 \
  --chunk-overlap=300 \
  --semantic

Process:

• Starts Docling Serve container (fixed port 5001)
• Starts pgvector container
• Clones Quarkus repository at version tag
• Finds and filters AsciiDoc guides (~251 guides)
• For each guide:
  • Extracts metadata from AsciiDoc headers
  • Fetches HTML from versioned quarkus.io URL
  • Converts to Markdown using Docling
  • Combines content with metadata
  • Ingests to pgvector with embeddings
• Dumps database to SQL
• Builds Docker image

Output: ghcr.io/quarkusio/chappie-ingestion-quarkus:3.15.0

Build time: ~15-20 minutes for full documentation set (250 guides)

Test with 5 guides:

java -jar target/quarkus-app/quarkus-run.jar bake-image \
  --quarkus-version=3.15.0 \
  --max-guides=5 \
  --semantic

Production build (all guides):

java -jar target/quarkus-app/quarkus-run.jar bake-image \
  --quarkus-version=3.15.0 \
  --chunk-size=1000 \
  --chunk-overlap=300 \
  --semantic

Using the image with chappie-server:

docker run -p 5432:5432 ghcr.io/quarkusio/chappie-ingestion-quarkus:3.15.0

Then configure chappie-server to connect to localhost:5432.

• Quarkus 3.30.8 - Application framework
• Quarkus Docling 1.2.2 - Document processing with layout awareness
• LangChain4J - RAG pipeline and pgvector integration
• BGE Small EN v15 - Embedding model (384 dimensions)
• Testcontainers - Container management (Docling + pgvector)
• Jib - Docker image building
• JGit - Git repository cloning

See src/main/resources/application.properties:

# Docling configuration (CDI injection with fixed port)
quarkus.docling.devservices.enabled=false
quarkus.docling.base-url=http://localhost:5001

# PGVector configuration
quarkus.langchain4j.pgvector.dimension=384
quarkus.langchain4j.pgvector.table=rag_documents
quarkus.langchain4j.pgvector.create-table=true
quarkus.langchain4j.pgvector.use-index=true
quarkus.langchain4j.pgvector.index-list-size=100

# PostgreSQL datasource (set by Testcontainers at runtime)
quarkus.datasource.db-kind=postgresql

# Logging
quarkus.log.level=INFO
quarkus.log.category."org.chappie".level=INFO

Note: The Docling Serve container is started programmatically by the bake-image command on port 5001.

Uses the same test suite from chappie-server:

• RagGoldenSetTest - 36 test cases for retrieval quality

Conclusion: The hybrid approach achieves 100% accuracy on the RAG golden set, improving upon the baseline.

✅ Complete and Production-Ready

• ✅ Hybrid architecture (AsciiDoc metadata + Docling HTML)
• ✅ CDI integration with fixed port (5001)
• ✅ Git repository cloning for version consistency
• ✅ Metadata extraction (topics, categories, extensions, summary)
• ✅ Versioned HTML fetching from quarkus.io
• ✅ Docling HTML → Markdown conversion
• ✅ Semantic chunking by headers
• ✅ Database dump and Docker image building
• ✅ 100% RAG golden set accuracy
• ✅ Documentation and examples

• Quarkus Docling Extension
• Build Agent-Ready RAG Systems with Quarkus and Docling
• LangChain4J Quarkus Integration
• APPROACH.md - Detailed implementation plan

Same as chappie-server project.