Information retrieval

Retrieval turns an answer you have to trust into an answer you can check.

A model can answer from what it already holds, or it can look things up first and show its work. This page shows the second mode on three real questions: the decision to retrieve, the evidence it pulled, the warnings attached, and the citations the answer is allowed to make.

These are fixed worked examples built from two local systems. The page runs no model and makes no live query. The evidence it shows is metadata about sources, never the source text itself.

Worked examples

Pick a question. Read both answers. Follow the evidence.

The question

Answer from memory

Answer with retrieval

What retrieval changed

The decision

Why retrieval happened

The evidence

Source packets, not source text

Each packet is metadata about a source: what it is, where it sits, why it matched, and any warning it carries. The page never prints the source text.

The citations

What the answer is allowed to cite

Technical fields for this example

Source and safety

What a public page can show, and what it holds back.

The page stays safe even if it is public. It names the systems and the public sources behind them. It does not carry the private index, the books, or anything that could leak.

Shown

Source titles, years, and public legal labels.
Retrieval-system metrics from build reports.
Packets that paraphrase the role of the evidence.
Warnings that mark local-only and current-source limits.

Withheld

Raw passages from any copyrighted book.
The private local index and any database file.
Access credentials, deployment keys, browser state, and logs.
Any live call from this page into a local index.

Appendix

Two systems, one pattern.

The examples borrow the shape of two local indexes. The page shows only counts, metadata, and paraphrased evidence.

Tax strategy retrieval

Index: SQLite FTS5
Sources: 72 indexed source and analysis files
Chunks: 25,536
Default output: metadata, tags, warnings, no passage
Role: precision mechanics and provenance, below the references

Living structure retrieval

Index: SQLite FTS5 with BM25
Layers: book text, analysis, reference
Chunks: about 4,600
Default output: title, year, layer, location, matched terms
Role: concept provenance, below the wholeness gate

What the words mean

RAG: Retrieval-augmented generation. The page uses plain language: the system looks things up before it answers, then shows what it found.
FTS5 and BM25: A local full-text search index and a ranking method that score chunks by matched terms and metadata, not by meaning vectors.
Packet: A cloud-safe unit of evidence: source title, year, layer, location, match, tags, and any warning. It is metadata, not the source text.
Currency guard: A rule that flags dated figures and routes any current number to a current official source before the answer can use it.

Design sources behind the layout

Tschichold, for form that follows the function of the text. Muller-Brockmann and Samara, for the grid as a system of relationships. Bringhurst, Hochuli, and Rutter, for measure, leading, and scale. Arnheim, Bertin, and Ware, for ordered hierarchy and color used as state. Cited as a source family; no passages are reproduced.