RAG System ├── Data Ingestion │ ├── PDF / DOCX / PPT / Notion / Web │ ├── Chunking strategies │ ├── Cleaning and deduplication │ ├── Metadata tagging │ └── Versioning ├── Embeddings │ ├── OpenAI / Cohere / BGE / Instructor │ ├── Vector normalization │ ├── Domain adaptation │ └── Re-embedding strategy ├── Vector Database │ ├── Pinecone / Weaviate / Qdrant / FAISS │ ├── Index types (HNSW, IVF, Flat) │ ├── Filtering and metadata search │ └── Sharding and scaling ├── Retrieval │ ├── Semantic search │ ├── Hybrid retrieval (BM25 + vector) │ ├── Re-ranking (cross-encoders) │ ├── Query rewriting │ └── Multi-hop retrieval ├── Prompt and Context Engineering │ ├── System, user and tool prompts │ ├── Context window management │ ├── Citation and grounding │ ├── Answer style control │ └── Safety and refusal rules ├── LLM Layer │ ├── GPT / Claude / Mistral / Llama │ ├── Temperature and decoding │ ├── Function calling and tools │ └── Cost and latency optimization ├── Generation │ ├── Context injection │ ├── Structured output (JSON) │ ├── Multi-step reasoning │ └── Tool-calling agents ├── Evaluation │ ├── Retrieval metrics (Recall@K, MRR) │ ├── Answer quality (faithfulness, relevance) │ ├── LLM-as-a-Judge │ └── Regression testing ├── Feedback Loop │ ├── User ratings │ ├── Click-through │ ├── Failure logging │ └── Data refresh triggers ├── Caching │ ├── Query caching │ ├── Embedding caching │ └── LLM response caching ├── Security and Guardrails │ ├── Prompt injection protection │ ├── Data isolation │ ├── PII filtering │ └── Tool permissioning ├── Observability │ ├── Token usage │ ├── Latency per step │ ├── Retrieval quality │ └── Hallucination tracking ├── Orchestration │ ├── LangChain / LlamaIndex / custom pipelines │ ├── Async flows │ └── Retry and fallback models ├── Infrastructure │ ├── API gateway │ ├── Vector DB hosting │ ├── Model routing │ └── Cost tracking └── Programming Language └── Python or TypeScript (start here) AI generated !!