**RAG is a pattern, not a framework feature.** The minimum viable pipeline is four steps:
1. Embed the user's query with an embedding model (OpenAI `text-embedding-3-small`, Voyage, Cohere, or a HuggingFace model).
2. Top-k nearest-neighbor search against a vector store (Pinecone, pgvector, Qdrant, Weaviate, Chroma, FAISS).
3. Format the retrieved chunks into the prompt as context.
4. Call the chat model with the augmented prompt.
With the OpenAI SDK and a vector client, this is roughly 80 lines of Python or TypeScript. No framework required.
**When does a framework earn its keep?** When you start needing things that are genuinely tedious to hand-roll:
- **Hybrid retrieval** (dense + BM25 fusion, optional rerank)
- **Routing** across multiple indices (one for docs, one for code, one for tickets)
- **Response synthesizers** that handle long retrieval contexts via map-reduce or refine patterns
- **Chunking experiments** with different splitters, overlap strategies, hierarchical chunks
- **Eval harnesses** that score retrieval recall and answer faithfulness
LangChain and LlamaIndex are great at these. They are not required to do RAG. Adopting them on day one is a common over-engineering trap.
**The distractors in this question** confuse RAG (the pattern) with one framework's marketing of the pattern. RAG existed before LangChain and exists outside it.
**RAG is a pattern, not a framework feature.** The pattern: embed query, top-k from a vector store, format the retrieved chunks into a prompt, call the model. Eighty lines around the OpenAI SDK (or any provider SDK) plus a vector client. That's a competitive pipeline.
The distractors in this question encode three common confusions. "Only LangChain has retrievers" mistakes a framework's API for the underlying capability, every vector store ships a nearest-neighbor query. "Only LlamaIndex's QueryEngine works" mistakes one abstraction for the pattern itself. "Requires a fine-tuned embedding model from the framework vendor" mistakes embedding choice for framework choice; off the shelf embeddings from OpenAI, Voyage, or Cohere work fine.
**When does a framework actually earn its keep?**
- **Hybrid retrieval** that fuses dense and BM25 with a reciprocal rank fusion or weighted score.
- **Multi-index routing**, a router LLM picks which retriever (docs / code / tickets) to query per request.
- **Response synthesis over long contexts**, map-reduce, refine, tree-summarize patterns when retrieval returns more chunks than fit in the model's context.
- **Chunking strategy experiments**, hierarchical chunking, sentence-window retrieval, parent-document retrieval.
- **Built-in tracing and eval hooks**, LangSmith, Phoenix, Langfuse integrations.
None of these is mandatory for a working RAG pipeline. They become tedious to hand-roll when the design grows past a single retriever and a single synthesis style. That's the right adoption trigger.
**The senior framing.** Adopt the framework when composition complexity makes hand-rolling tedious, not when you read "RAG" in a requirements doc.
Quick reference
TL;DR
RAG is a pattern, not a framework feature: embed query, top-k from a vector store, format retrieved chunks into the prompt, call the model. About 80 lines around any provider SDK.
Imagine asking your grandma a cooking question while sliding her a few pages from a recipe book you grabbed off the shelf. She reads the pages, then answers using what she just saw. That is the whole idea. You do not need a fancy kitchen in a box subscription to do this. You need a recipe book, a way to flip to the right pages, and grandma. Anyone selling you a special toolkit for it is selling convenience, not the basic move. The fancy toolkit becomes worth buying only when you start juggling many cookbooks, mixing in handwritten notes, or running a busy diner. For one cook at one counter with one book, the simple version works fine and stays out of your way.
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example.
Adopting a framework on day one because the requirements doc says "RAG". The pipeline is 80 lines; the framework adds dependency weight, lock-in, and learning cost before any complexity has justified it.
A team standardized on a non-framework stack (raw provider SDK + custom retrieval) where adding a framework fragments the codebase. Mixed-paradigm RAG codebases are harder to reason about than uniformly hand-rolled or uniformly framework-driven ones. Pick one and commit.
A prototype to validate whether RAG even helps for the use case. Hand-rolling first surfaces real failure modes (bad chunks, off-topic retrieval, hallucinated synthesis) that framework defaults sometimes paper over. Once you know what works, then choose whether the framework is worth adopting.
Quick scan · 2–4 min
Everything important, quickly.
State that RAG is a pattern (4 steps: embed, retrieve, augment, generate), name the 80-line shape around any provider SDK, list the framework-earning capabilities (hybrid, routing, response synthesizers, chunking experiments, eval/obs), and close with the adoption-trigger framing.
Real-world examples
Real products, models, and research that use this idea.
1Klarna's customer-support assistant runs a custom RAG pipeline on top of the OpenAI SDK with a self-managed vector store, no LangChain in the hot path.
2Notion's Q&A feature is reported to run on a hand-rolled retrieval layer rather than a third-party framework, with framework code reserved for offline experimentation.
3Perplexity's production search-augmented generation uses its own retrieval and ranking layer over multiple search backends rather than a single framework's abstractions.
4LangChain and LlamaIndex are widely used for RAG, the point is that they're a choice, not a requirement; many production RAG systems exist outside both.
Interviewer follow-ups
What an interviewer would ask next. Try answering before peeking at the approach.
QWhat's the concrete trigger that should move a team from 80-line hand-rolled RAG to a framework?
Sign in or go Pro to see the approach for this follow-up.
QHow would you measure whether your RAG pipeline is actually working without adopting a framework?
Sign in or go Pro to see the approach for this follow-up.
QWhen you DO adopt LlamaIndex or LangChain for RAG, what specific primitives are you really after?
Sign in or go Pro to see the approach for this follow-up.
Don't say this
Red flags and common mistakes that signal junior thinking. Click to expand.
Most common mistake
Adopting LangChain or LlamaIndex on day one because "we need RAG". The pipeline is 80 lines; the framework earns its keep when composition complexity arrives, not before.