Unopened items
IKEA Canada says new, unopened products can generally be returned within 365 days with proof of purchase for a refund to the original payment method.
Read the official pageShyam Sarma / Index
Contact02 / AI Systems LabLive RAG demo
Everyday Help Desk
Index finds the right piece of information.
Ask a question about an IKEA return, an iPhone backup, or Spotify downloads. See the source notes broken into pieces, watch Index find the most relevant ones, then ask it to write an answer with citations. Each step is yours to inspect.
01 / Source notes2026-09-13 snapshot
Technical trace How this run is actually computed
Corpus: 12 curated chunks from three official help topics, snapshot 2026-09-13. Corpus embeddings are precomputed with text-embedding-3-small at 64 dimensions and bundled server-side.
At retrieval time, the question is embedded with the same model. Server-side cosine similarity ranks all 12 vectors. The connected three-axis PCA graph is visualization only; neither the drawn edges nor a shortest-path algorithm determine retrieval.
Generation receives the top four chunks. It uses GPT-5.6 Luna to write a short answer with chunk-ID citations. The API re-runs retrieval server-side before generation instead of trusting evidence sent by the browser.
How it works
How Index is built.
The demo exposes the retrieval pipeline, while keeping embeddings, API keys, and ranking on the server. Each visual stage corresponds to a real processing step.
- 01Curate & chunk
Short source-backed notes are split into labeled, inspectable passages.
- 02Embed once
Each passage gets a 64-number OpenAI embedding stored with the server build.
- 03Retrieve live
A new question is embedded, then compared with every passage using cosine similarity.
- 04Generate
Luna sees only the highest-ranked notes and produces a cited answer.
Engineering notes
Architecture, tools, and tradeoffs.
What each piece does, and why it was included in this demo.
Small corpus, no vector DBTwelve vectors fit in a versioned server bundle.
For this fixed educational corpus, a database would add operational weight without improving the demonstration. A larger, frequently updated corpus would justify a dedicated index.
The 3D networkPCA visualizes; full vectors decide.
The nodes come from a principal-component projection of the actual embeddings. The thin edges create a connected visual scaffold, and the glowing query paths show the highest-ranked results. Search itself compares all 64 dimensions; it does not traverse the displayed edges.
Evidence boundarySource text is data, not an instruction.
The generation call is built from a server-side top-four retrieval, and the model is told to answer only from those notes and cite their IDs. The notes are curated summaries, not live copies of the official sites.
Cost and abuse controlsDaily per-visitor and global limits.
The public endpoints use durable Redis-backed limits and reject requests when the limits or required configuration are unavailable. Retrieve and Generate have separate budgets.