There are currently two types of fulltext searching: an SQL-based word index and a file scanner. They each have their advantages and drawbacks.
The word index is very fast to search and is currently used for the find-as-you-type quicksearch. However, indexing files takes some time, so we should probably offer a preference to turn it off ("Index attachment content for quicksearch" or something). There's also an issue with Chinese characters (which are indexed by character rather than word, since there are no spaces to go by, so a search for a word with common characters could produce erroneous results). The quicksearch doesn't use a left-bound index (since that would probably upset German speakers searching for "musik" in "nachtmusik," though I don't know for sure how they think of words) but still seems pretty fast.
* Note: There will be a potentially long delay when you start Firefox with this revision as it builds a fulltext word index of your existing items. We obviously need a notification/option for this. *
The file scanner, used in the Attachment Content condition of the search dialog, offers phrase searching as well as regex support (both case-sensitive and not, and defaulting to multiline). It doesn't require an index, though it should probably be optimized to use the word index, if available, for narrowing the results when not in regex mode. (It does only scan files that pass all the other search conditions, which speeds it up considerably for multi-condition searches, and skips non-text files unless instructed otherwise, but it's still relatively slow.)
Both convert HTML to text before searching (with the exception of the binary file scanning mode).
There are some issues with which files get indexed and which don't that we can't do much about and that will probably confuse users immensely. Dan C. suggested some sort of indicator (say, a green dot) to show which files are indexed.
Also added (very ugly) charset detection (anybody want to figure out getCharsetFromString(str)?), a setTimeout() replacement in the XPCOM service, an arrayToHash() method, and a new header to timedtextarea.xml, since it's really not copyright CHNM (it's really just a few lines off from the toolkit timed-textbox binding--I tried to change it to extend timed-textbox and just ignore Return keypress events so that we didn't need to duplicate the Mozilla code, but timed-textbox's reliance on html:input instead of html:textarea made things rather difficult).
To do:
- Pref/buttons to disable/clear/rebuild fulltext index
- Hidden prefs to set maximum file size to index/scan
- Don't index words of fewer than 3 non-Asian characters
- MRU cache for saved searches
- Use word index if available to narrow search scope of fulltext scanner
- Cache attachment info methods
- Show content excerpt in search results (at least in advanced search window, when it exists)
- Notification window (a la scraping) to show when indexing
- Indicator of indexed status
- Context menu option to index
- Indicator that a file scanning search is in progress, if possible
- Find other ways to make it index the NYT front page in under 10 seconds
- Probably fix lots of bugs, which you will likely start telling me about...now.
Note that there's no code for user types and fields yet -- just the schema (actually there's a tiny bit of code in the item type manager, since we'll probably use some of the same methods for managing user types, but not much)
Templates for primary item types are currently only used by the item type manager to make creating new types easier and to prevent the removal of fields from an item type that are associated with its template item type -- the fields are all still recorded in itemTypeFields, since they might have different orders or default visibility settings from their templates
This lets us add tables to user.sql without writing migration steps for them yet still have the ability to change existing user tables and migrate data if necessary.
Also added _getDropCommands() to do a regex on the SQL file and create the DROP TABLE|INDEX steps necessary to use the DB_REBUILD flag without DROP commands in the SQL file itself, before I realized that it probably made the most sense to just delete the SQL file and storage directory. Changed _initializeSchema() to do that instead. Leaving _getDropCommands() in, in case there's ever a need for it.
This should make development much easier, as we can, for example, add 80 item types without having to write transition steps
Pretty sure this won't delete anyone's data. Might want to test that theory, though.
