Select all components that are required to build a minimal end to end RAG system (the absolute basics, not nice to haves like rerankers or eval harnesses).
Minimal RAG has four required parts: chunker, embedding model, vector index, and generator LLM. Rerankers and fine-tuning are not on that list; one is a quality lift, the other leaves the pattern.
Picture the smallest RAG system you can build. You have a folder of documents and a user with a question. To go from one to the other, you need four things. Something to chop the documents into bite size pieces, so retrieval can be precise. Something to turn each piece (and the question) into a number coordinate, so 'similar meaning' becomes a measurable distance. Something to store those coordinates and find the closest ones fast. And something to read the closest pieces and write an answer for the user. Take any one of those four away and the system stops working. Rerankers, fine-tunes, eval harnesses, and clever query rewriters are all good ideas later, but none of them are in the minimum kit. And fine-tuning is not even part of the pattern: it changes the model, which RAG is specifically designed to avoid.
Concept explanation~2 min read
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Concept explanation~2 min read
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
This is the most common architectural question in any RAG interview, and it filters cleanly. Candidates who can list the four required components and explain why each one is non-optional show they have built a RAG system or at least walked through one carefully. Candidates who include rerankers or fine-tuning in the minimum show they have read marketing material about RAG but not built one.
The useful framing is that RAG is a pattern with a small architectural commitment. Four moving parts, two flows (index-time and query-time), one defining property (knowledge in the retrieval system, not in the weights). Everything else, from rerankers to query expansion to hybrid retrieval to eval harnesses to observability, is optional quality lift that production systems add but minimal systems do not need.
This deep dive walks through each of the four required components in detail, traces the index-time and query-time flows that compose them, then explains why the reranker is correctly placed outside the minimum and why fine-tuning belongs to a different pattern entirely.
The four required components
A minimal RAG system has four moving parts, and each one closes a specific gap that the system would not work without.
The chunker runs at indexing time. It takes long source documents and splits them into smaller pieces, typically 200 to 800 tokens. Without chunking, the alternative is to embed each whole document as a single vector, which produces a blurry semantic average that matches many queries weakly and none of them sharply. Chunking is what gives retrieval precision. Even though chunking is preprocessing, it is fundamental to the pattern, not optional.
The embedding model turns text into vectors. At indexing time it embeds each chunk. At query time it embeds the user query. The same model must be used on both sides, because cosine similarity is only meaningful inside a single learned geometry. Common 2026 choices include OpenAI's text-embedding-3-large (3072-dim), Voyage Embed v3 (asymmetric variants), Cohere Embed v4 (multilingual), and BGE-M3 (open weight). Without an embedding model, semantic similarity has no numerical representation to compare against.
The vector index stores the embeddings and supports fast approximate nearest neighbor search. Common choices are Pinecone, Qdrant, Weaviate, Milvus, pgvector on Postgres, and Chroma. The index uses algorithms like HNSW or IVF to find the top-k closest vectors in tens of milliseconds, instead of the seconds a brute force scan would take. Without the index, retrieval is O(N) and breaks past a few thousand chunks.
The generator LLM is the final piece. It takes the user query plus the retrieved chunks and produces the natural language answer. Common 2026 choices are Claude Opus 4.7, GPT-5.5, Gemini 3.1 Pro, and Llama 4. The generator stays frozen across queries; nothing about it is fine-tuned for the RAG corpus. Its job is to read the retrieved evidence and synthesize an answer that grounds in that evidence.
Situations where this technique stops working.
2–4 min · Everything important, quickly.
Real products, models, and research that use this idea.
- LangChain's basic RAG template shipping exactly these four parts: TextSplitter, OpenAIEmbeddings, Chroma vector store, and a ChatOpenAI generator.
- LlamaIndex's quickstart pipeline using SimpleNodeParser, OpenAI embeddings, an in-process vector store, and Claude Opus 4.7 as the generator.
What an interviewer would ask next. Try answering before peeking at the approach.
QHow would you extend the minimal four-part system to make it production-ready?
Add a reranker, hybrid retrieval (BM25 plus vector), evaluation harness with Ragas, observability and tracing, caching, access control on the index.
Red flags & common mistakes
The phrases that signal junior thinking. Click to expand.
Red flags & common mistakes
The phrases that signal junior thinking. Click to expand.
Listing a reranker or a fine-tuned model as part of the minimum architecture. Rerankers are a quality lift on top; fine-tuning leaves the RAG pattern entirely because RAG is built around a frozen generator.
60 second bullets to scan on the way to the call.
Primary sources. Browse if you want the original framing.
Same topic, related formats. Practice these next.