[ChinaYiqun]

What changed

• introduces a minimal ChatService abstraction so the RAG pipeline can be tested without network calls;
• keeps OpenAIChatService as the default OpenAI-compatible implementation;
• changes NaiveRAG from a demo-shaped chain into an actual RAG pipeline;
• calculates the query embedding once instead of recalculating it during sorting;
• batches chunk embedding through EmbeddingService#getEmbeddings;
• validates pipeline order, blank input, missing chunks, empty vectors, and dimension mismatches;
• selects configurable Top-K chunks after ranking;
• builds a grounded prompt containing the original question and retrieved passages;
• instructs the model to treat retrieved text as untrusted data and to cite passages as [1], [2], etc.;
• retains parsering() as a deprecated compatibility alias for the correctly named parsing() method;
• replaces the commented demo test with deterministic, offline unit tests.

Root cause

The previous NaiveRAG.LLMChat() ignored both the user query and retrieved chunks. It always sent the fixed prompt 1+1 = ? to the model. The embedding phase also computed the query vector and discarded it, forcing the sorting phase to make a second provider call.

User impact

The minimal pipeline now performs the behavior its name promises:

parse -> chunk -> embed -> rank -> Top-K context -> grounded generation

The full constructor supports dependency injection for alternative splitters, embedding models, and chat providers while the original two-argument constructor remains available.

Validation

CI runs all Maven tests. The new tests verify that:

• the query embedding is called once;
• chunk embeddings are batched once;
• the nearest passage becomes the selected context;
• the original user question reaches the model;
• the selected passage reaches the model;
• a passage outside Top-K does not reach the model;
• sorting before embedding fails explicitly.