dojo.provide("dojox.data.WikipediaStore"); dojo.require("dojo.io.script"); dojo.require("dojox.rpc.Service"); dojo.require("dojox.data.ServiceStore"); dojo.experimental("dojox.data.WikipediaStore"); dojo.declare("dojox.data.WikipediaStore", dojox.data.ServiceStore,{ // summary: // Initializer for the Wikipedia data store interface. // description: // The WikipediaStore is a data store interface to Wikipedia, using the // Wikipedia SMD spec from dojox.rpc. It currently is useful only for // finding articles that contain some particular text or grabbing single // articles by full name; no wildcards or other filtering are supported. // example: // | var store = new dojox.data.WikipediaStore(); // | store.fetch({ // | query: {title:"Dojo Toolkit"}, // | onItem: function(item){ // | dojo.byId("somediv").innerHTML = item.text["*"]; // | } // | }); constructor: function(options){ if(options && options.service){ this.service = options.service; }else{ var svc = new dojox.rpc.Service(dojo.moduleUrl("dojox.rpc.SMDLibrary", "wikipedia.smd")); this.service = svc.query; } this.idAttribute = this.labelAttribute = "title"; }, fetch: function(/* object */ request){ // summary: // Fetch a page or some partially-loaded search results from // Wikipedia. Note that there isn't a way to sort data coming // in from the API, so we just ignore the *sort* parameter. // example: // Loading a page: // | store.fetch({ // | query: {title:"Dojo Toolkit"}, // | // define your handlers here // | }); // example: // Searching for pages containing "dojo": // | store.fetch({ // | query: { // | action: "query", // | text: "dojo" // | }, // | // define your handlers here // | }); // example: // Searching for the next 50 pages containing "dojo": // | store.fetch({ // | query: { // | action: "query", // | text: "dojo", // | start: 10, // | count: 50 // max 500; will be capped if necessary // | }, // | // define your handlers here // | }); var rq = dojo.mixin({}, request.query); if(rq && (!rq.action || rq.action === "parse")){ // default to a single page fetch rq.action = "parse"; rq.page = rq.title; delete rq.title; }else if(rq.action === "query"){ // perform a full text search on page content rq.list = "search"; rq.srwhat = "text"; rq.srsearch = rq.text; if(request.start){ rq.sroffset = request.start-1; } if(request.count){ rq.srlimit = request.count >= 500 ? 500 : request.count; } delete rq.text; } request.query = rq; return this.inherited(arguments); }, _processResults: function(results, def){ if(results.parse){ // loading a complete page results.parse.title = dojo.queryToObject(def.ioArgs.url.split("?")[1]).page; results = [results.parse]; }else if(results.query && results.query.search){ // loading some search results; all we have here is page titles, // so we mark our items as incomplete results = results.query.search; var _thisStore = this; for(var i in results){ results[i]._loadObject = function(callback){ _thisStore.fetch({ query: { action:"parse", title:this.title }, onItem: callback }); delete this._loadObject; } } } return this.inherited(arguments); } });