org.llm4s.javaapi.rag

Retrieval-augmented generation from Java: ingest documents, then ask questions that a model answers from the chunks of them that match, with those chunks returned as the answer's sources. It is llm4s-rag's pipeline - chunking, embedding, a vector and keyword (hybrid) search and the answer prompt - with no Scala type in sight.

try (JRag rag = JRag.builder()
       .embeddings(Llm4s.createDefaultEmbeddingClient().get())
       .llm(Llm4s.createDefaultClient().get())
       .build()
       .get()) {
   rag.ingest(Path.of("docs")).get();
   JRagAnswer answer = rag.ask("How do I configure a provider?").get();
   System.out.println(answer.answer());
   answer.sources().forEach(s -> System.out.println("  " + s.documentId() + " (" + s.score() + ")"));
}

Obtain one from JRag.builder. Every call blocks the calling thread and never throws: a failure - of the embedding provider, the model, a file or an argument - comes back inside the LlmResult, whose getError().getKind() says which, as for JLlmClient and JEmbeddingClient. Interrupting the blocked thread returns a failed result of kind CANCELLED, with the interrupt flag still set.

Different documents may be ingested from several threads at once; ingesting the same document id from two threads at once is not supported. close() releases the store - the SQLite connections, or the memory of an in-memory store

• and not the clients it was built with, which stay the caller's to close.