Data & Storage

AI Chat stores everything inside your application: structured data in your App's database via OrmLite, and files under App_Data. There is no external service holding your conversation history.

Database tables​

Tables are created on startup when AutoInitSchema is true (the default), using the host's IDbConnectionFactory - or a named connection if you'd rather keep chat data separate:

services.AddPlugin(new ChatFeature {
    NamedConnection = "chat",
    AutoInitSchema = true,
});

ChatThread​

One conversation. Complex fields are stored as raw JSON strings so the wire shape matches the OpenAI format exactly, and are only parsed at the DTO boundary.

StreamingMessage is deliberately separate from Messages: a failed or abandoned stream can never damage the durable conversation, and checkpointing writes one small column rather than rewriting the whole thread.

ChatRequest​

Per-completion accounting behind the Analytics dashboards - user, thread, model, provider, duration, token counts, prices, cost, finish reason, and Error/StackTrace when a completion fails.

ChatMedia​

The generated and uploaded media catalog - name, type, prompt, model, cost, seed, dimensions, size, duration, aspect ratio, content hash, reactions, caption, tags, rating and publish state. Written by the gallery extension from a cache-saved filter. See Voice & Media.

Gemini tables​

The gemini extension owns its own document catalog tables, created the same way. See Gemini RAG, Search & Analytics.

Querying​

They're ordinary OrmLite tables:

using var db = dbFactory.Open();

var recent = db.Select<ChatThread>(db.From<ChatThread>()
    .Where(x => x.User == userName)
    .OrderByDescending(x => x.UpdatedAt)
    .Take(20));

AdminQueryChatRequests additionally exposes ChatRequest as an AutoQuery API for admins.

File storage​

Rooted at App_Data/chat by default, overridable with AppDataPath:

Per-user layout​

With RequireAuth = false everything runs as the default user, so all of the above lives under user/default/.

Path resolution is guarded: any relative path that would escape App_Data/chat throws UnauthorizedAccessException.

Static project exports​

The global user/default/share_static/config.json configures static folder sharing. By default, exports live outside App_Data at <WebContentDirectory>/p/<user>/<project-folder>/, typically wwwroot/p/.... A custom Directory can select another export root. The UI abbreviates this root to ~/; it does not display the full destination path.

Project staticPublication metadata remains in the user's project configuration, separately from the remote publication link. Folder exports can be served by a static server without AI.Chat running. Include the export directory in deployment or backup if you need to retain the served copies. See Publishing.

The content-addressed cache​

Cache writes fire the cache_saved filters, which is how the gallery records media and how an App can hook uploads:

ctx.RegisterCacheSavedFilter(saved => {
    Log.LogInformation("Cached {Url} ({Size} bytes)", saved.Url, saved.Size);
});

Cached files are served at {RoutePrefix}/~cache/{path} to authenticated users.

Seeded config files​

llms.json is seeded only when missing and preserves user settings across restarts. Provider catalogs (providers.json and providers-extra.json) refresh from bundled defaults on startup unless their filenames are listed in PreserveConfigs. Preserved catalog files are seeded when missing.

services.AddPlugin(new ChatFeature {
    PreserveConfigs = ["providers-extra.json"],
});

Alternatively, bypass llms.json entirely by setting ChatFeature.Config in code. See Providers & Models.

PDF storage​

PdfFeature uses a separate root, App_Data/pdf by default:

See Rendering PDFs.

Backup and deployment​

Back up together:

• Your App's database (threads, requests, media, Gemini catalog)
• App_Data/chat - config, cache and every user's workspace
• App_Data/pdf - including .published.json and .versions

A database backup on its own is not sufficient: media rows reference cache files by hash, and Gemini document rows reference cached bytes.

Deployment notes:

Retention​

AI Chat doesn't expire data on your behalf, so retention policy is yours to implement:

// example: delete threads untouched for a year
using var db = dbFactory.Open();
var cutoff = DateTime.UtcNow.AddYears(-1);
var ids = db.Column<long>(db.From<ChatThread>()
    .Where(x => x.UpdatedAt < cutoff).Select(x => x.Id));

db.Delete<ChatRequest>(x => Sql.In(x.ThreadId, ids));
db.Delete<ChatThread>(x => Sql.In(x.Id, ids));

Deleting a user's folder under App_Data/chat/user/{user} removes their projects, profiles, skills, PDF workspace and preferences.

Decision, subscription, and creation state​

Use a single owning host process per App_Data root for Decision Studio and rotating subscription credentials. Stop hosts before moving profiles and preserve hidden files and pending receipts. Back up the database and App_Data together. A downgrade requires restoring the matching pre-upgrade backup rather than handing migrated state to an older binary.