0. Data Preprocessing Chunk Optimization Embedding Model Optimization (DAPT, TAPT, etc .. ) Make MetaData RAFT 1. Query Analysis (= Pre-Retriever) Multi Query Decomposition HyDE 부정어가 섞여있는 query의 경우 부정적인 쿼리를 filtering해서 가져올 수 있는지? (ex, 산은 좋고, 바다는 싫어 -> 바다에 대한 relevant docs는 걸러내야 함) 자연어 query를 넣었을 때와 한국어 토크나이징을 수행한 query의 relevant docs 추출 성능비교 2. Query Routing (= Pre-Retriever) Logical Routing Semantic Routing 3. Retriever & Indexing Lexical : BM25 Lexcial : BM42 (advanced) Semantic : DPR Semantic : PDR Ensemble : Hybrid Search Indexing : FAISS Indexing : HNSW RDB + Vector (= Docker + ParadeDB) 4. Filtering (= Post-Retriever) Meta-Data Filtering Similarity threshold cut-off Passage Filtering : Refine Passage Filtering : Compressor 5. Re-Ranking (= Post-Retriever) API Base LM Base LLM Base Language Specific Model Base 6. Prompting Long-Context Reorder (-> Lost in the middle) Zero-shot & Few-shot Prompt-Chaining CoT ToT TDB ReAct Reflextion 7. Generation & Inference LLM Base Generation Data type(fp16, fp32, bp16, nf16, etc..) vllm Ollama DSPy 8. Hallucination Control CRAG Self-RAG etc.. 9. Evaluation Evaluation (each RAG step) Relevance (hitrate) Groundtruth 10. Methology Memory moudle (from Modular RAG) Tabular Data Reasoning RE-Questioning RAG Module Implementation
