Dainty Trading
Home / Projects / ChatVault

ChatVault

One searchable archive for every conversation you’ve had with every model. Claude, ChatGPT, Gemini, Grok, Perplexity — all in one place, indexed by what you actually said and what they actually answered.

Category  AI · Knowledge management Stack  Python · Postgres · pgvector Sources  Claude · ChatGPT · Gemini · Grok · Perplexity Status  In final testing

The problem

If you use multiple LLMs — and most serious users now do — your knowledge is scattered. The brilliant breakdown of a contract clause is in Claude. The code review that actually caught the bug is in ChatGPT. The one Gemini answer you keep meaning to come back to is somewhere on a different device. None of them talk to each other; none of them are searchable across providers; and none of them survive the moment you change subscriptions.

What we’re building

ChatVault is a single archive that ingests conversations from every major LLM provider — Claude, ChatGPT, Gemini, Grok, and Perplexity — through their export formats and APIs. Each conversation is stored in Postgres with full-text and semantic indexes (Postgres full-text + pgvector), plus a normalised schema for who-said-what, when, with which model, and at what cost.

Search works two ways: keyword (fast, exact, finds the Slack-paste of a regex) and semantic (slower, fuzzy, finds “that thing about negotiation tactics from a few weeks ago” without the right keywords). Results show the surrounding turns so you get the context, not just the snippet.

The AI angle

The interesting work is the recall layer. Embeddings catch the gist; re-ranking with Claude promotes the result that actually answers your question. We log the click-through — which result the user opened, whether they marked it useful — and use it to tune ranking over time. Over a few weeks of use the archive becomes meaningfully tailored to how you actually search.

How it’ll be used

  • Power users of multiple models who want one place that doesn’t care which provider produced the answer.
  • Researchers and writers who use AI for thinking partners and want every prior thread retrievable.
  • Engineers revisiting code reviews, prompt iterations, and debugging conversations across tools.

Where we are

Ingest is working for Claude exports, ChatGPT exports, Gemini history, and Perplexity threads. Grok is the latest addition. Semantic recall is in private testing with a re-rank step. Next milestone is a small browser extension that captures conversations as you have them, instead of waiting on export-and-import.