Skip to main content

Document folders

Most cases come with a pile of files: filings, contracts, exported emails, transcripts, notes. Rather than turning every one into a Document node, attach the folder to the board and let DeepState index it. The assistant can then search the whole set by meaning, quote the passages it finds, and tell you which file and page they came from.

Attaching a folder​

Open a board, then go to Settings → Documents and click Attach folder…. Indexing starts right away and runs in the background; the progress bar shows the current file. Attach as many folders as you like. Files that change are re-indexed on the next scan (⟳), files you delete are dropped, and files that were already indexed are skipped, so a re-scan of a large folder is quick.

Supported: PDF (with a text layer), Word (.docx), text, Markdown, HTML, .eml email exports (headers and body, multipart aware), CSV and JSON. Files over 50 MB and PDFs without a text layer (scans) are listed as skipped; OCR for folder documents is not applied yet.

What indexing does​

Each file is split into passages of roughly a paragraph or two, with a little overlap so a sentence that straddles a boundary is still found. Every passage is embedded with nomic-embed-text and stored in the board's database next to a full-text index. Searching combines the two: passages that are close in meaning and passages that contain the exact words, fused so that a hit found both ways ranks first.

Embedding runs on your machine. If Ollama is running with the nomic-embed-text model pulled, DeepState uses it (and your GPU). Otherwise it uses the same model built in, on the CPU; the model (about 130 MB) is downloaded once into DeepState's data folder. Indexing happens on a background thread, so the app stays responsive; expect about a minute per few hundred pages on a laptop CPU.

Nothing about the folder leaves your computer. The index is derived data: it is not included in sync, and a colleague who opens the same board indexes their own copy of the folder.

Using it​

  • Ask the assistant. "What do the documents say about the Harbourview sale?" The search_documents tool returns the best passages with file and page; read_document reads on from a hit. The assistant cites what it quotes.
  • Try it yourself. The search box at the bottom of the Documents settings tab runs the same search, so you can see what a question will surface before you ask.
  • Mentions. After indexing, DeepState looks for the names (and aliases) of the board's actors and organizations in every file. An actor's panel lists the documents that mention them; the assistant sees the same list when it reads a document.

Removing a folder​

The trash icon removes the folder and everything indexed from it. The files themselves are never touched.