Overview — Project 2: AskDocs

Project 1 converted one frozen page into one artifact. Project 2 builds AskDocs, a CLI that answers questions from a directory of Markdown and text files. You will build the small honest version of RAG: chunk, embed, retrieve, assemble context, generate, and cite.

Run the finished first slice

cd rust
cargo run -p askr -- --mock --rag askr/sample-docs \
  "What does Rust use instead of a garbage collector?"

The reference crate uses a deterministic hashing embedder and mock language model when no key is available. The output is not supposed to be intelligent; it proves which chunks were selected and what context crossed into the model boundary.

The four days

Day	Build	Rust learned by needing it	AI system learned by observing it
6	directory ingestion and chunks	iterators, strings, paths, ownership	chunking changes retrievability
7	offline embeddings and top-k	traits, slices, floats, lifetimes	similarity is a ranking signal
8	grounded prompt and citations	borrowed results, formatting, errors	retrieval is not proof of support
9	streaming plus frozen questions	async, callbacks, byte buffers, tests	UX and quality need separate measures

Product pipeline

docs ─▶ chunks ─▶ embeddings ─▶ vector store
                                  │
question ─▶ query embedding ─▶ top-k ─▶ context manifest ─▶ model ─▶ cited answer

The companion has two levels:

• rust/askr is the small project you can understand end to end;
• rust/rust-ai-engineering contains the hardened retrieval, evidence, snapshot, and provider contracts that later projects reuse.

Do not start with a vector database. First make ranking visible with a Vec, cosine similarity, three documents, and assertions you can calculate by hand.

Start with Day 6 — Ingest and Chunk →.