Boox 1.5: from local chat to file analysis and web research
Boox v1.5.0 is out. It adds agent mode for text-file analysis and web research alongside ordinary chat. You can run it on your own machine with Docker and Ollama, or choose a supported cloud model.
The useful change is being able to continue a task. Upload a CSV, ask for a total, then ask how the answer was calculated. The follow-up keeps the file available. You can also enable web research, inspect the activity, and follow the sources in the answer.
The demo uses fictional data and local Granite 3B. Waiting time is shortened.
A small example
The demo uses a fictional totals.csv with three amounts: 5, 7 and 9. In Files-only agent mode, ask for the total. The expected result is 21. A follow-up asks the agent to explain the addition. No tool names or special “do not search the web” wording are needed.
For research, clear the retained files and enable Web. Ask the question in ordinary language and request sources. The activity panel shows the work the agent performed and can be collapsed while you read the answer.
What runs locally
With an Ollama model, model inference runs locally. Conversation history and text attachments are stored in SQLite on the machine running Boox.
Web research still contacts external search providers. Cloud-model selection sends task input, including attached content, to that provider. Files-only turns off web tools; it does not make a cloud model local.
Legacy search remains the default. This release also adds optional Parallel MCP search and public-page reading, configured in the backend. Provider access and any costs are separate from Boox.
The limits matter
Agent mode accepts up to four text, Markdown, CSV, JSON or code files, with 24,000 characters combined. PDF and Office documents are not supported. It does not edit files or execute code. Conversation memory and web-tool calls are bounded, so long conversations can lose earlier detail.
Answers depend on the model. Source links give you something to inspect; they do not establish that every claim is correct. Release checks with the compact default, Granite 3B, covered chat, file analysis, follow-ups and web research. One published-image web check exceeded the three-minute model timeout before the full suite passed on retry. Slow local inference remains a practical constraint.
Boox is intended for a trusted, single-user setup. Its APIs are unauthenticated, so restrict network access rather than exposing the installation publicly.
Try v1.5.0
Download the versioned Compose installer from the v1.5.0 release and follow the setup instructions. It selects dumanoj/boox:1.5.0 and downloads the default model on first start. Existing users should back up their database and configuration before upgrading; the release notes include backup and rollback steps.
The code is on GitHub. If you try it, feedback on file follow-ups, source quality and small-model behavior would be especially useful. Include the model and steps to reproduce when reporting an issue, and keep private files and credentials out of reports.