Build your own RAG pipeline

What you'll build

You'll build the retrieval and context-assembly half of a RAG pipeline in TypeScript, end to end: chunk a frozen SQuAD corpus three different ways, embed and index the chunks, retrieve top-k with MMR de-duplication, and pack a token-budgeted context block with [doc i] citations. Because SQuAD answers are verbatim spans, you prove the context actually contains the answer with an exact string check — and prove your citations point at the right chunk — with no LLM judge anywhere.

• Three real chunkers — fixed-size with token overlap, the LangChain recursive splitter, and a LlamaIndex-style sentence-window splitter — emitting token-exact chunk boundaries
• A retrieval stage over ONNX MiniLM embeddings with MMR de-duplication, plus token-budget context assembly that formats [doc i] citation markers and tracks sources
• A retrieve→assemble→generate→cite loop where the generator is a deterministic cassette keyed by the exact assembled prompt
• A CLI — theinference rag ask "<question>" — that prints the exact context block, the grounded answer, its citations, and a flag proving the gold answer was present

What you'll learn

• How a RAG pipeline really decomposes — and why most failures are retrieval, not the model: the answer never made the top-k, or the citation pointed at the wrong chunk
• How chunk size and chunk strategy move retrieval quality, measured in a real ablation rather than guessed
• How to evaluate retrieval deterministically with recall@k, MRR, and answer-span hit-rate@k over gold chunk ids — no LLM grading itself
• How to enforce citation integrity by checking the cited chunk actually contains the verbatim answer span

Prerequisites

• Comfortable reading and running TypeScript
• Familiar with embeddings and cosine similarity — vectors that represent text, compared by dot product (the vector-db / semantic-search basics)
• No prior RAG experience required — we build chunking, retrieval, and the metrics from scratch