Melville · Moby-Dick (demo)

start with a question

select a result to read

Shodō 書道

a map to the library of your AI universe

Borges's Library of Babel contained every possible book and no way to find any of them — total information, total uselessness. Your AI conversation history is a smaller version of the same problem: vast, complete, and effectively inaccessible. Shodō is the catalog those wanderers in the hexagons prayed for.

Shodō ingests your exported AI conversations — Claude, ChatGPT, and others — and indexes them into a local corpus with full semantic search. Every exchange is embedded, chunked, and threaded. Conversations are titled and tagged by an LLM enrichment pass.

Search works at the level of meaning, not just keywords. Ask a question; Shodō retrieves the conversations where that question was answered.

sageframe.net ↗ built by Andrew Todd Marcus

HOW TO SEARCH

The one rule: name a thing, not an operation. Match the moment's language — what was on screen when the conversation happened, not your later description of it. "Money anxiety and the pressure to finish" works. "conversations about the writing process" does not.

six ways searches fail

temporal "first," "earliest" — results by relevance, not time. Use exact‑language mode + date order.

lexical exact phrase trace → use exact‑language mode with quotes: "wicked book"

vocabulary generic term collides with another domain → anchor with a discriminating term.

compositional multi-step operation → decompose into a query sequence, each narrowing the last.

over-specified asking retrieval to verify a condition → loosen to what's retrievable, judge downstream.

diffuse signal behavioral pattern, not a topic → query the language of the moment, not the pattern name.

Broad-terms tax: common words (book, writing, work) pull off-domain hits. Pair them with something domain-specific to dominate the match.

Reading results: scores below ~0.12 are noise-adjacent. The same conversation appearing multiple times in top-k is a stronger signal than any single score. Raise top k when you want a full set, not a reading sample.