Seafile 14 adds an AI chat that answers over your own libraries
On 14 October 2026, the Seafile team opened testing for version 14.0, headlined by AI Chat — an assistant that explores your libraries through their metadata — and SeaSearch, a full-text and vector search engine now set as the default. Try it on a throwaway instance first: self-hosted does not mean the AI runs locally.
14 October 2026. The Seafile team opens testing for version 14.0 on its forum. SeaSearch 1.0 becomes the default search engine, with full-text and vector search. 14 October 2026, again. AI Chat now answers questions about a library from its metadata. Why it matters: a self-hosted file server, until now a mute tree of folders, becomes able to answer a question — provided you understand, before pointing it at real documents, where the data goes.
AI Chat: an assistant that reads metadata, not bytes
The headline feature is AI Chat. Its stated job is to “find and understand content in your libraries” through conversation. Instead of descending folder by folder, you ask — “find the files related to this project,” “what is under this folder,” “analyze the file distribution in this library” — and the assistant answers by cross-referencing metadata.
The exact scope matters, and it is more modest than a “chatbot that understands every byte on your server.” AI Chat works on metadata: file names, paths, sizes, AI-generated summaries, and other properties. It can flag potentially duplicate files or summarize what a folder contains. This is conversational search over metadata, not a reading of every document’s raw content. The nuance matters: privacy claims of “your data stays home” hinge precisely on what is sent to a model and what is not.
The comparison with rivals helps place the offering. Nextcloud has added semantic search through its AI assistant, Paperless-ngx indexes scanned documents, and Immich finds photos by description. Seafile stands apart by working first on library metadata and organization — a conservative choice that limits what passes through a model, but does not replace the per-document content extraction that specialized tools offer. The angle is less “find a PDF” than “understand what is in a shared space.”
SeaSearch: full-text and vector search by default
The second piece is SeaSearch, which becomes the default search engine at version 1.0. It combines full-text search with vector search, the latter drawing on AI-generated summaries to find documents by meaning rather than exact wording. Regular search lets you pick the scope — the current library, the current folder, or all libraries. AI Chat operates at the library level.
Around that core, 14.0 modernizes the rest: customizable file properties (surface AI summaries and tags in the list), a new data table view, a metadata server component for extended properties, and real-time interface updates when others add or remove files. Share links gain descriptions, webhooks can report a modified library, and format support broadens — TIFF, .mkv, .m4v, EPUB thumbnails, and .csv via OnlyOffice.
One dependency is easy to miss: vector search only pays off once the AI summaries it relies on actually exist. A fresh library has no summaries until you generate them, so a brand-new Seafile 14 instance starts with full-text search working and vector search essentially empty. The quality of AI Chat’s answers therefore scales with the completeness of that metadata layer — which is another way of saying the feature rewards patience and punishes a same-day production cutover. It also means the first hours of evaluation tell you little: you have to let the summarization jobs run before conversational search has anything meaningful to retrieve. For anyone coming from a folder-tree workflow, that is the main mental shift — the search gets better the longer the library has been indexed, not the longer you have been using it.
Standing up a test instance
The test comes together in minutes with Docker, without touching an existing installation. The project ships dedicated images for evaluating search and chat in isolation:
services:
seafile:
image: seafileltd/seafile-mc:14.0.0-preview
ports:
- "8080:80"
volumes:
- ./seafile-data:/sharedOn that throwaway instance, load a small non-sensitive library, ask questions whose answers you already know, and compare AI Chat against regular search. The point is not to validate the feature on its best result, but on its edge cases: near-identical file names, deliberately unhelpful folder names, documents you know by heart. That is where metadata search proves itself — or breaks.
The trap: self-hosted does not mean local AI
This is the point announcements never hammer hard enough. Hosting Seafile yourself does not guarantee that all AI operations stay on your machine. Vector search and AI summaries depend on a model — and depending on the edition, the configuration, and the chosen model provider, that model can run locally or be called remotely. The 14.0 documentation explicitly mentions a Seafile AI extension to redeploy, with model settings configured in the server’s environment file. “Self-hosted” describes where the file server lives, not where inference runs.
The operational consequence is simple to state: before connecting real documents, check the edition, the model configuration, and the data flow. A Community Edition instance with no external API key sends nothing out; an instance wired to a LLM provider sends at minimum metadata, and possibly content, depending on what you enable. The reflex of “it’s self-hosted, so it’s private” is exactly what a CISO should break.
Under the hood: new components to master
14.0 is not just an AI layer bolted onto the existing stack: it adds components you will have to deploy and size. A dedicated metadata server now manages extended file properties; a notification server lets the interface update in real time when others add or remove files. SeaSearch, for its part, is configured in the server’s environment file: search engine choice, document content indexing, and connection settings.
For a Docker installation, upgrading means redeploying the Seafile AI extension, removing the old obsolete settings before reconnecting the new one. That is the kind of migration you rehearse on a copy, not on the production server on a Friday night. The project’s message is clear: the version is ready for testing, not for production. The 14.0 upgrade notes explicitly list configuration changes to absorb before any cutover.
Verdict
If you already run Seafile in production, touch nothing: 14.0 is a testing release, not general availability, and a premature upgrade on an active file server has never improved a Monday. If you want to evaluate conversational search, stand up a throwaway instance with a small, non-sensitive library, ask questions whose answers you already know, and compare AI Chat against regular search rather than judging on the best result. Before connecting any real data, verify the edition, the model, and the data flow — and keep a rollback plan with verified backups outside the experiment. Seafile 14 does not reinvent file storage; it finally makes the question worth asking — “what is actually in all these libraries?” — provided you do not mistake the answer for a privacy promise.