Version-Aware RAG Demo

Try it: filter OFF vs ON

Your question is embedded in your browser with the same model as the Python system (all-MiniLM-L6-v2). The page then searches the 17 stored chunks twice: once by similarity alone, once with the version filter. The first search downloads the model (~90 MB, cached afterwards).

Model not loaded yet; it loads on the first search.

Filter OFF: similarity only

—

Filter ON: `superseded = False`

—

The 21 frozen test cases

Written and committed before the fix. These are the measured results from the Python system, retrieval only, top 4. Click a question to run it live above.

ID	Type	Question	OFF	ON	Superseded retrieved OFF → ON

How it works

Every stage existed in the class lab except the version filter. The metadata comes from a separate manifest.json, so the embedded text is exactly what a reader sees.

16 runbooks→load→+ version metadata→chunk→embed (MiniLM)→Chroma

question→embed→VERSION FILTER→top-4 by similarity→LLM answer (Python version only)

On this page, the stored chunk vectors come from the Python pipeline, and questions are embedded in your browser with transformers.js (fp32). Checked before publishing: for all 21 test questions, filter OFF and ON, the browser returns the same top 4 in the same order as the Python system (42/42).

The corpus: 8 operations runbooks, two versions each

Example runbooks for a fictional company's messaging platform. Each pair shares wording and differs in its facts. Old versions carry no "superseded" banner; only the metadata knows.