Vector DB

What: Each example card becomes a 64-number vector from hashed content words. A query is embedded the same way. Cosine distance picks the nearest neighbors. No Ollama.
Why: This is the floor under RAG. If retrieval is messy, generation will sound sure and still be wrong. A local embedder lets you see the geometry even when Llama is not running.
What you are seeing: Flow on the left lights as each step finishes. AWS Architecture on the right is that step’s box — click the title for the stacked twin. On GitHub Pages the hash runs in the browser — no localhost API. Click a finished step for cards, dimensions, or neighbors.

Static GitHub Pages demo — sample questions replay recorded runs with paced animations. Vector DB hashes in this tab and does not call localhost. Clone the repo and make dev for live labs.

This example hashes in your browser. It does not call the local API or Llama. Embedder: hashed_tokens · 64 dimensions.

Flow

AWS Architecture

START
accept the query
END
return neighbors

index

keep vectors in RAM

11 vectors sit in process memory (LocalVectorIndex). Restart wipes them. Nothing is written to disk.

Raw update

{
  "embedder": "hashed_tokens",
  "dim": 64,
  "ready": true,
  "count": 11
}