AI agents for your filesystem: semantic search and auto-organization

The problem with file search

You know the file exists. You wrote it last Tuesday. It was about the Q3 budget. But what was it called? report.xlsx? q3-numbers-final-v2.xlsx? Traditional file search forces you to remember names. That's broken.

DoubleSpace already has FTS5 full-text search — you can find files by their contents. But text matching isn't understanding. Searching for "budget" won't find a file that discusses "quarterly financial projections" even though they mean the same thing.

Semantic search with Vectorize

Cloudflare's Vectorize is a vector database that runs on the edge. When a file is uploaded to DoubleSpace, our indexer queue extracts the text content, generates an embedding via Workers AI, and stores it in Vectorize. Queries are embedded the same way and matched against the index using cosine similarity.

The result: search by meaning. Ask "files about revenue projections" and get back documents that discuss financials, even if they never use the word "revenue."

# Traditional search — exact match
$ dblspc search "budget"
  → budget-2026.xlsx
  → budget-template.docx

# Semantic search — meaning match
$ dblspc search "quarterly financial projections" --semantic
  → Q3-planning-deck.pptx
  → revenue-forecast-h2.xlsx
  → board-update-july.pdf
  → budget-2026.xlsx

The AI file agent

Search is just the beginning. We're building an AI agent that sits on top of your filesystem and can answer questions, summarize documents, and organize files — all using Workers AI models running on the edge.

Ask your files

Instead of searching and then reading, just ask:

$ dblspc ask "What was our revenue target for Q3?"
  Based on Q3-planning-deck.pptx and revenue-forecast-h2.xlsx:
  The Q3 revenue target was $2.4M, up 15% from Q2.
  Source files: /work/finance/Q3-planning-deck.pptx (slide 4)

Auto-tagging

When files are uploaded, the agent can automatically classify them — invoices, contracts, meeting notes, code, design files. Tags are stored as metadata and are searchable.

$ dblspc ls /inbox --show-tags
  contract-acme-2026.pdf      [contract] [legal] [acme-corp]
  meeting-notes-apr-3.md       [meeting] [engineering]
  screenshot-2026-04-05.png   [screenshot] [ui]
  invoice-cloudflare-mar.pdf  [invoice] [infrastructure]

Smart organization

The agent can suggest folder structures based on content patterns, move files into appropriate directories, and flag duplicates or near-duplicates.

Architecture

The AI layer adds two Cloudflare services to the stack:

• Workers AI — runs embedding models (bge-base-en-v1.5) and LLMs (Llama) on the edge for text extraction, embedding generation, question answering, and classification
• Vectorize — stores embeddings for semantic similarity search, scoped per workspace

The indexer queue already runs on every upload. Adding AI is an extension of that pipeline: extract text → generate embedding → store in Vectorize → classify with LLM → write tags. All asynchronous. The upload response is instant.

Privacy model

Because DoubleSpace runs on your Cloudflare account, your files and embeddings never leave your infrastructure. Workers AI processes data in-region and doesn't retain inputs. There's no third-party AI provider with a copy of your documents.

This is a fundamentally different privacy model from Google Drive's "AI features" or Dropbox Dash — those systems process your files on their servers. With DoubleSpace, the AI runs on your edge.

Status and roadmap

DoubleSpace currently ships with FTS5 full-text search (Tier 1 — filename matching, Tier 2 — content search). The AI features described here are Tier 3, currently in development:

• ✓ Full-text search (shipped)
• ✓ Automatic text extraction on upload (shipped)
• ◐ Vectorize integration (in progress)
• ○ Semantic search CLI flag (planned)
• ○ dblspc ask command (planned)
• ○ Auto-tagging pipeline (planned)
• ○ Smart organization agent (planned)

Follow the project on GitHub or check the blog for updates.