A terminal for asking questions against markdown libraries. Hybrid search (BM25 + pgvector) in one local PGlite database, via a local model.

Install

Node.js 20+ has to be on PATH. Ask and search need an OpenAI-compatible /v1/chat/completions and /v1/embeddings server. Ollama exposes those at http://127.0.0.1:11434/v1.

npx @papodaca/hideout

From a clone:

npm install && npm run build && node dist/cli.js

That also puts hideout on PATH if you npm link, or you can run npx @papodaca/hideout after publish.

hideout
hideout extract book.pdf markdown/book
hideout index
hideout search "your query"

extract splits a PDF by its bookmark outline. If the file has no bookmarks, it writes one markdown file per page.

Type a question at the prompt. /search prints raw hits. /config edits the chat model, API urls, headers, and source directories. /help lists commands.

Paths

Config: $XDG_CONFIG_HOME/hideout/config.json (default ~/.config/hideout/config.json)

Index (PGlite + pgvector), set toggles, and prompt history: $XDG_DATA_HOME/hideout/ (default ~/.local/share/hideout/). First launch applies schema migrations. Prompt history lives in the PGlite database; the TUI loads the last 50 lines so up-arrow works.

First run writes a config if none exists. If ./markdown or ./docs is present, those get added as sources. Otherwise point sources at one or more markdown directories. Subfolders become document sets you can toggle with /sets.

{
  "chat_model": "ornith-1.5:9b",
  "embed_model": "nomic-embed-text",
  "chat_url": "http://127.0.0.1:11434/v1",
  "embed_url": "http://127.0.0.1:11434/v1",
  "banner": true,
  "sources": [
    "/path/to/your/markdown"
  ],
  "chat_headers": {
    "Authorization": "Bearer ${API_KEY}"
  },
  "embed_headers": {
    "Authorization": "Bearer ${API_KEY}"
  }
}

Header values expand ${ENV_VAR} from the environment.