Chat search

Purpose

Chat search scans chat titles and message text. It builds an index per workspace. Desktop builds persist that index outside the workspace folder.

Indexing idea

We build the index from chat app trees. Each entry stores a thread title, messages, searchText, and updatedAt. Tokenization is simple and runs locally.

Storage

We persist the index per workspace. On desktop, store it in app data, outside the workspace directory. Path: app.getPath("userData")/search-index/<spaceId>/chat-index.json. This avoids sync conflicts in local-first workspaces. Outside desktop, we keep the index in memory. It rebuilds when you open chat search.

Format (desktop storage):

    {
  "version": 1,
  "updatedAt": 1710000000000,
  "entries": [
    {
      "threadId": "app-tree-id",
      "title": "Chat title",
      "messages": ["message text"],
      "searchText": "chat title message text",
      "updatedAt": 1710000000000
    }
  ]
}
  

Query flow

• Tokenize query.
• Match threads that contain all tokens.
• Score by term frequency with title boost.
• Sort by score, then updatedAt. Queries run in the renderer using the in-memory entries. The renderer debounces queries on typing.

Update flow

• Load the persisted index if it exists.
• Compare app tree updatedAt to cached entries.
• Rebuild only changed threads.
• Keep cached messages when unchanged.
• Regenerate searchText when needed.

Limitations

• No document/file search.
• No fuzzy match or stemming.
• Large workspaces still load each app tree to check updatedAt.

Next iterations

• Use a richer index format (JSONL or SQLite).
• Add change observers to refresh without opening search.
• Add file/document search in the same UI.
• Improve ranking (recency, field weighting).

Desktop backend

Electron stores the index in the main process. The UI loads/saves it through IPC.

IPC handlers:

• sila:chat-search:load-index
• sila:chat-search:save-index

Renderer bridge: window.desktopSearch.