AI / 01 · Interactive lab

Retrieval, step by step.

Give it a document and a question, and watch retrieval-augmented generation step by step: each passage placed by its embedding, every passage scored against the question, the top K kept, and an answer that cites the ones it used.

Load the sample document Use your own document

Loading the lab…

How it's built.

The stage is a view over a real pipeline. In the cloud, the embeddings and the answer come from hosted models through my gateway; locally, all of it runs in your browser. Every number on the stage comes from this run on your document; token counts are estimates, about four characters a token.

Limits, stated plainly

Parsing and chunking always run in your browser. Cloud mode sends passage text and your question through my gateway, which stores nothing, to a hosted model provider whose own retention policy applies. Local mode sends nothing after a one-time model download; a document indexed in the cloud is re-embedded in your browser first.
Three dimensions keep only the broad shape of a 384- or 1024-dimension embedding. The map shows that shape; the ranking, the lens and the K boundary use the full vectors.
Retrieval is an exhaustive in-memory scan: right for hundreds of passages, not for millions.
Text-based files only: PDF text, Word (.docx) and plain text. Scanned PDFs would need OCR.

01Chunk: Split on paragraphs. A paragraph over 800 characters is split by sentence with 15% overlap, and very short ones are merged. 800 characters fits the local model's 256-token window. Runs in your browser.
02Embed: Cloud: the gateway's embedding model, 1024 dimensions. Local: all-MiniLM-L6-v2 (23 MB, quantized) in a Web Worker, 384 dimensions, mean-pooled and normalized. The map is a PCA projection of these vectors.
03Retrieve: Exact cosine similarity against every passage, top K kept. The view then re-centres on the question and puts each passage at a distance set by 1 − cosine, rescaled to fit.
04Assemble: Each passage wrapped in its own <evidence> tag with delimiters escaped, and the system prompt tells the model to treat evidence as data, not instructions. This lowers prompt-injection risk; it doesn't remove it.
05Generate: A hosted model through my gateway, SmolLM2 locally on WebGPU, or the prompt copied to another assistant.

AI and backend engineer. I build tools like this to make system behaviour visible.

GitHub profile

AI / 01 · Interactive lab

Retrieval, step by step.

Load the sample document Use your own document

Loading the lab…

How it's built.

Limits, stated plainly

Parsing and chunking always run in your browser. Cloud mode sends passage text and your question through my gateway, which stores nothing, to a hosted model provider whose own retention policy applies. Local mode sends nothing after a one-time model download; a document indexed in the cloud is re-embedded in your browser first.
Three dimensions keep only the broad shape of a 384- or 1024-dimension embedding. The map shows that shape; the ranking, the lens and the K boundary use the full vectors.
Retrieval is an exhaustive in-memory scan: right for hundreds of passages, not for millions.
Text-based files only: PDF text, Word (.docx) and plain text. Scanned PDFs would need OCR.

01Chunk: Split on paragraphs. A paragraph over 800 characters is split by sentence with 15% overlap, and very short ones are merged. 800 characters fits the local model's 256-token window. Runs in your browser.
02Embed: Cloud: the gateway's embedding model, 1024 dimensions. Local: all-MiniLM-L6-v2 (23 MB, quantized) in a Web Worker, 384 dimensions, mean-pooled and normalized. The map is a PCA projection of these vectors.
03Retrieve: Exact cosine similarity against every passage, top K kept. The view then re-centres on the question and puts each passage at a distance set by 1 − cosine, rescaled to fit.
04Assemble: Each passage wrapped in its own <evidence> tag with delimiters escaped, and the system prompt tells the model to treat evidence as data, not instructions. This lowers prompt-injection risk; it doesn't remove it.
05Generate: A hosted model through my gateway, SmolLM2 locally on WebGPU, or the prompt copied to another assistant.

AI and backend engineer. I build tools like this to make system behaviour visible.

GitHub profile

Retrieval, step by step.

The lab

How it's built.

Retrieval, step by step.

The lab

How it's built.