Knot

Getting Started

First, run the development server:

npm run dev
# or
yarn dev
# or
pnpm dev
# or
bun dev

RAG

Quiz generation can use retrieved course chunks from Snowflake when the dashboard sends a course id. If these env vars are set, the quiz API will embed the unit/topic query, retrieve top-k chunks, and ask Gemini to base at least 80% of answers on that material. If they are missing or retrieval fails, the app falls back to topic-only generation.

• SNOWFLAKE_HOST – Snowflake account host (e.g. abc12345.snowflakecomputing.com)
• SNOWFLAKE_TOKEN – Bearer token for REST API
• SNOWFLAKE_TOKEN_TYPE – Optional (default PROGRAMMATIC_ACCESS_TOKEN)
• SNOWFLAKE_DATABASE – Database name (default KNOT)
• SNOWFLAKE_WAREHOUSE – Warehouse for queries
• SNOWFLAKE_ROLE – Optional role

Same DB/schema as the ingestion pipeline (see ingestion/README.md).

Testing that RAG is used

1. Response debug – Every quiz response includes _debug: { ragUsed: boolean, chunkCount: number, query?: string }. In the browser: DevTools → Network → trigger a quiz → select the generate request → Response tab. If ragUsed is true and chunkCount ≥ 1, RAG retrieval ran and that many chunks were injected into the prompt.
2. Chunk info in response – When RAG is used, the response includes a sources array: each item has chunk_id, document_id, document_title, course_name, module_name, score, and text. The client receives it via generateQuiz() (returns { questions, sources? }); the Quiz component keeps sources in state for citations or "view source" later.
3. Server logs – With npm run dev, when RAG is used you’ll see a log like: [Quiz RAG] courseId=... unitId=... chunks=8 query="Unit: ...".
4. Compare with/without – Use a course that has ingested chunks. Open a unit and start a quiz (dashboard sends courseId). Then call the API manually without courseId (e.g. omit it in a curl body); the first run should show ragUsed: true, the second ragUsed: false.
5. Content check – If your course material has distinctive terms or definitions, RAG-generated questions should reflect them; non-RAG questions may be more generic.