ragleap-rag (Java port)

version 0.10.0 · Java 21

A pure-Java port of ragleap-rag for JVM teams: a live-verified RAG pipeline on Postgres/pgvector, streaming generation with provider fallback, and six vector backends (Milvus and Pinecone are code-complete, not live-verified).

<dependency>
  <groupId>io.github.antonyrag</groupId>
  <artifactId>ragleap-rag</artifactId>
  <version>0.10.0</version>
</dependency>

Highlights

• Core pipeline: HikariCP connection pooling, 7 embedding providers, pgvector dense, sparse and hybrid (RRF) search
• Generation: streaming, structured output, provider fallback, vision captioning
• Postgres-backed conversation memory and query rewriting: contextual rewrite, HyDE, multi-query with Reciprocal Rank Fusion
• Five vector backends: pgvector, FaissBackend (pure Java, no JNI), QdrantBackend, WeaviateBackend, MilvusBackend
• Qdrant and Weaviate backends are live-verified against real instances; MilvusBackend is explicitly not live-verified, and its release notes say so
• 8 standalone utility modules, including chunking, sanitization, guardrails, cost tracking and evaluation scoring

Tech stack

Requires Java 21. Dependencies: jtokkit, json-schema-validator, HikariCP, postgresql JDBC driver, sqlite-jdbc (metadata sidecar for the non-pgvector backends).

Testing

278 @Test methods across 26 test files, counted directly from the repo. Qdrant and Weaviate tests run against real instances and skip cleanly when none is reachable; Milvus tests use a local stub server because no real instance was available.