Context Management Project — RAG over a document corpus

A starter that extends the Foundations chat with retrieval-augmented generation. You'll implement the indexing pipeline and wire retrieval + citations into the chat backend.

What's in here

rag-starter/
├── documents/                  20 real Wikipedia articles (Apollo missions)
├── indexer.py                  walk docs → chunk → embed → store
├── backend/
│   ├── app.py                  extended chat with RAG stubs
│   └── requirements.txt
├── frontend/                   React UI (Foundations chat + Sources display)
└── .env.example

The corpus is 20 Wikipedia articles (plain-text extracts):

• 15 Apollo missions: 1, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17
• 5 related: Apollo program (overview), Saturn V, Lunar Module, Command/Service Module, Mission Control Center

Total ~850 KB of text across the 20 files. After chunking (~1000 chars each) you'll have several hundred chunks. Real enough that some questions answer crisply and others surface the seams.

Setup

# from this directory
python3.11 -m venv .venv
source .venv/bin/activate
pip install -r backend/requirements.txt

cp .env.example .env
# set ANTHROPIC_API_KEY

The first time you run anything that imports sentence-transformers, it will download the embedding model (multilingual, ~470 MB). One-time only.

Your job

1. Implement chunking and build the index

Open indexer.py. There's a TODO for chunk_text(). Implement it (see the Context Management lecture slide for one working version). Then:

python indexer.py
# Indexing documents from documents/
#   01-overview.md: 3 chunks
#   02-streaks.md: 4 chunks
#   ...
# ✓ Indexed N chunks → index.pkl

2. Wire retrieval into the chat backend

Open backend/app.py. There are TODOs for:

• Updating SYSTEM_PROMPT with citation rules
• Calling search(user_message, INDEX, k=5) to get top chunks
• Formatting them as a numbered context block
• Building user_content with CONTEXT: + QUESTION:

The citation parser is already wired — it returns the source filenames the model cited, which the frontend already displays under each answer.

3. Run it

# Terminal 1 — backend
cd backend
python app.py

# Terminal 2 — frontend
cd frontend
npm install
npm run dev

Open http://localhost:5173. Ask questions about the Apollo program. You should see:

• An answer that draws on the indexed docs
• A Sources: line citing which files were used

Try these (graded easy → hard):

• "What was the cause of the Apollo 1 fire?" — single-doc, factual.
• "Which Apollo missions landed on the Moon?" — cross-cutting, enumeration.
• "Compare the moonwalk durations of Apollo 11 and Apollo 17." — cross-doc, comparison.
• "List Apollo missions that used the Saturn V rocket." — cross-doc, requires reasoning over the corpus.
• "What is the Artemis program?" — out-of-corpus; the system should say it doesn't know rather than hallucinate.

4. Report

Pick 5 test questions that probe the corpus from different angles. For each:

• The question
• The answer the system gave
• Whether the cited sources are correct (open the file, verify)
• A judgment: did the system answer well, weakly, or hallucinate?

Then write up 2 strengths and 2 weaknesses of your implementation with the worked examples as evidence.

What to present

• Your chunking choice (size, overlap, boundary rule) and why
• Your system prompt's citation rules
• One question that works cleanly, with the right citations
• One question that fails — wrong answer, missing citation, or hallucinated source
• What you would change (chunking? retrieval? prompt?) to fix the failure

Alternative project

Want to RAG over your own corpus (your notes, a docs site you've cloned, a code repo's READMEs)? Replace the contents of documents/ and re-run python indexer.py. The rest of the pipeline works unchanged.