Beispiel #1
0
        private List <BookSearchModel> SearchBookContent(string searchWords)
        {
            List <BookSearchModel> bookSearchModelList = new List <BookSearchModel>();
            //1.对搜索条件进行分词
            Analyzer    analyzer    = new PanGuAnalyzer();
            TokenStream tokenStream = analyzer.TokenStream("", new StringReader(searchWords));

            Lucene.Net.Analysis.Token token = null;
            string indexPath = @"D:\lucenedir";
            //string kw = "面向对象";//对用户输入的搜索条件进行拆分。
            FSDirectory   directory = FSDirectory.Open(new DirectoryInfo(indexPath), new NoLockFactory());
            IndexReader   reader    = IndexReader.Open(directory, true);
            IndexSearcher searcher  = new IndexSearcher(reader);
            //搜索条件
            PhraseQuery query = new PhraseQuery();

            //foreach (string word in kw.Split(' '))//先用空格,让用户去分词,空格分隔的就是词“计算机   专业”
            //{
            //    query.Add(new Term("body", word));
            //}
            //query.Add(new Term("body","语言"));--可以添加查询条件,两者是add关系.顺序没有关系.
            // query.Add(new Term("body", "大学生"));
            while ((token = tokenStream.Next()) != null)
            {
                query.Add(new Term("body", token.TermText()));
            }
            // query.Add(new Term("body", kw));//body中含有kw的文章
            query.SetSlop(100);//多个查询条件的词之间的最大距离.在文章中相隔太远 也就无意义.(例如 “大学生”这个查询条件和"简历"这个查询条件之间如果间隔的词太多也就没有意义了。)
            //TopScoreDocCollector是盛放查询结果的容器
            TopScoreDocCollector collector = TopScoreDocCollector.create(1000, true);

            searcher.Search(query, null, collector);                                    //根据query查询条件进行查询,查询结果放入collector容器
            ScoreDoc[] docs = collector.TopDocs(0, collector.GetTotalHits()).scoreDocs; //得到所有查询结果中的文档,GetTotalHits():表示总条数   TopDocs(300, 20);//表示得到300(从300开始),到320(结束)的文档内容.
            //可以用来实现分页功能
            for (int i = 0; i < docs.Length; i++)
            {
                //
                //搜索ScoreDoc[]只能获得文档的id,这样不会把查询结果的Document一次性加载到内存中。降低了内存压力,需要获得文档的详细内容的时候通过searcher.Doc来根据文档id来获得文档的详细内容对象Document.
                int             docId       = docs[i].doc;         //得到查询结果文档的id(Lucene内部分配的id)
                Document        doc         = searcher.Doc(docId); //找到文档id对应的文档详细信息
                BookSearchModel searchModel = new BookSearchModel();
                searchModel.Id                = int.Parse(doc.Get("ID"));
                searchModel.Title             = doc.Get("title");
                searchModel.ContenDescription = SearchWordHighlight.CreateHightLight(searchWords, doc.Get("body"));
                //this.listBox1.Items.Add(doc.Get("number") + "\n");// 取出放进字段的值
                //this.listBox1.Items.Add(doc.Get("body") + "\n");
                //this.listBox1.Items.Add("-----------------------\n");
                bookSearchModelList.Add(searchModel);
            }
            //将搜索的此插入词库之中
            SearchDetails entity = new SearchDetails()
            {
                Id = Guid.NewGuid(), KeyWords = searchWords, SearchDateTime = DateTime.Now
            };

            SearchDetailsService.AddEntity(entity);
            return(bookSearchModelList);
        }
Beispiel #2
0
        public ActionResult Search(string wd = "", int page = 1, int size = 15)
        {
            var nul = new List <PostOutputDto>();

            ViewBag.Elapsed  = 0;
            ViewBag.Total    = 0;
            ViewBag.PageSize = size;
            ViewBag.Keyword  = wd;
            if (Regex.Match(wd ?? "", CommonHelper.BanRegex + "|" + CommonHelper.ModRegex).Length > 0)
            {
                return(RedirectToAction("Search"));
            }

            string ip  = HttpContext.Connection.RemoteIpAddress.MapToIPv4().ToString();
            string key = "Search:" + ip;

            if (CacheManager.Exists(key))
            {
                var hotSearches = RedisHelper.Get <List <KeywordsRankOutputDto> >("SearchRank:Week").Take(10).ToList();
                ViewBag.hotSearches = hotSearches;
                ViewBag.ErrorMsg    = "10秒内只能搜索1次!";
                return(View(nul));
            }

            wd = wd?.Trim().Replace("+", " ");
            if (!string.IsNullOrWhiteSpace(wd) && !wd.Contains("锟斤拷"))
            {
                if (!HttpContext.Session.TryGetValue("search:" + wd, out _) && !HttpContext.Request.IsRobot())
                {
                    SearchDetailsService.AddEntity(new SearchDetails
                    {
                        KeyWords   = wd,
                        SearchTime = DateTime.Now,
                        IP         = HttpContext.Connection.RemoteIpAddress.MapToIPv4().ToString()
                    });
                    SearchDetailsService.SaveChanges();
                    HttpContext.Session.Set("search:" + wd, wd.ToByteArray());
                }

                var posts = PostService.SearchPage(page, size, wd);
                ViewBag.Elapsed = posts.Elapsed;
                ViewBag.Total   = posts.Total;
                if (posts.Total > 1)
                {
                    CacheManager.AddOrUpdate(key, wd, s => wd);
                    CacheManager.Expire(key, TimeSpan.FromSeconds(10));
                }

                ViewBag.hotSearches = new List <KeywordsRankOutputDto>();
                return(View(posts.Results));
            }

            ViewBag.hotSearches = RedisHelper.Get <List <KeywordsRankOutputDto> >("SearchRank:Week").Take(10).ToList();
            return(View(nul));
        }
        public ActionResult Search(string wd = "", int page = 1, int size = 15)
        {
            var nul = new List <PostDto>();

            ViewBag.Elapsed  = 0;
            ViewBag.Total    = 0;
            ViewBag.PageSize = size;
            ViewBag.Keyword  = wd;
            string ip  = ClientIP;
            string key = "Search:" + ip;

            if (CacheManager.Exists(key) && CacheManager.Get(key) != wd)
            {
                var hotSearches = RedisHelper.Get <List <KeywordsRank> >("SearchRank:Week").Take(10).ToList();
                ViewBag.hotSearches = hotSearches;
                ViewBag.ErrorMsg    = "10秒内只能搜索1次!";
                return(View(nul));
            }

            wd = wd?.Trim().Replace("+", " ");
            if (!string.IsNullOrWhiteSpace(wd) && !wd.Contains("锟斤拷"))
            {
                if (!HttpContext.Session.TryGetValue("search:" + wd, out _) && !HttpContext.Request.IsRobot())
                {
                    SearchDetailsService.AddEntity(new SearchDetails
                    {
                        KeyWords   = wd,
                        SearchTime = DateTime.Now,
                        IP         = ClientIP
                    });
                    SearchDetailsService.SaveChanges();
                    HttpContext.Session.Set("search:" + wd, wd.ToByteArray());
                }

                var posts = PostService.SearchPage(page, size, wd);
                ViewBag.Elapsed = posts.Elapsed;
                ViewBag.Total   = posts.Total;
                if (posts.Total > 1)
                {
                    CacheManager.AddOrUpdate(key, wd, s => wd);
                    CacheManager.Expire(key, TimeSpan.FromSeconds(10));
                }

                ViewBag.Ads         = AdsService.GetByWeightedPrice(AdvertiseType.PostList);
                ViewBag.hotSearches = new List <KeywordsRank>();
                return(View(posts.Results));
            }

            ViewBag.hotSearches = RedisHelper.Get <List <KeywordsRank> >("SearchRank:Week").Take(10).ToList();
            ViewData["page"]    = new Pagination(page, size);
            return(View(nul));
        }
        //private void CreateSearchContent()
        //{
        //    string indexPath = @"D:\WebSites\lucenedir";//注意和磁盘上文件夹的大小写一致,否则会报错。将创建的分词内容放在该目录下。
        //    FSDirectory directory = FSDirectory.Open(new DirectoryInfo(indexPath), new NativeFSLockFactory());//指定索引文件(打开索引目录) FS指的是就是FileSystem
        //    bool isUpdate = IndexReader.IndexExists(directory);//IndexReader:对索引进行读取的类。该语句的作用:判断索引库文件夹是否存在以及索引特征文件是否存在。
        //    if (isUpdate)
        //    {
        //        //同时只能有一段代码对索引库进行写操作。当使用IndexWriter打开directory时会自动对索引库文件上锁。
        //        //如果索引目录被锁定(比如索引过程中程序异常退出),则首先解锁(提示一下:如果我现在正在写着已经加锁了,但是还没有写完,这时候又来一个请求,那么不就解锁了吗?这个问题后面会解决)
        //        if (IndexWriter.IsLocked(directory))
        //        {
        //            IndexWriter.Unlock(directory);
        //        }
        //    }
        //    IndexWriter writer = new IndexWriter(directory, new PanGuAnalyzer(), !isUpdate, Lucene.Net.Index.IndexWriter.MaxFieldLength.UNLIMITED);//向索引库中写索引。这时在这里加锁。
        //    List<Books> list = BooksService.LoadEntities(t => true).ToList();
        //    foreach (var booksModel in list)
        //    {
        //        Document document = new Document();//表示一篇文档。
        //        //Field.Store.YES:表示是否存储原值。只有当Field.Store.YES在后面才能用doc.Get("number")取出值来.Field.Index. NOT_ANALYZED:不进行分词保存
        //        document.Add(new Field("Id", booksModel.Id.ToString(), Field.Store.YES, Field.Index.NOT_ANALYZED));

        //        //Field.Index. ANALYZED:进行分词保存:也就是要进行全文的字段要设置分词 保存(因为要进行模糊查询)

        //        //Lucene.Net.Documents.Field.TermVector.WITH_POSITIONS_OFFSETS:不仅保存分词还保存分词的距离。
        //        document.Add(new Field("Title", booksModel.Title, Field.Store.YES, Field.Index.ANALYZED, Lucene.Net.Documents.Field.TermVector.WITH_POSITIONS_OFFSETS));
        //        document.Add(new Field("ContentDescription", booksModel.ContentDescription, Field.Store.YES, Field.Index.ANALYZED, Lucene.Net.Documents.Field.TermVector.WITH_POSITIONS_OFFSETS));
        //        writer.AddDocument(document);
        //    }
        //    writer.Dispose();//会自动解锁。
        //    directory.Dispose();//不要忘了Close,否则索引结果搜不到
        //}

        private List <ViewModelContent> ShowSearchContent()
        {
            string        indexPath = @"D:\WebSites\lucenedir";
            List <string> list      = Common.WebCommon.PanGuSplitWord(Request["txtSearch"]);//对用户输入的搜索条件进行拆分。
            FSDirectory   directory = FSDirectory.Open(new DirectoryInfo(indexPath), new NoLockFactory());
            IndexReader   reader    = IndexReader.Open(directory, true);
            IndexSearcher searcher  = new IndexSearcher(reader);
            //搜索条件
            PhraseQuery query = new PhraseQuery();

            foreach (string word in list)//先用空格,让用户去分词,空格分隔的就是词“计算机   专业”
            {
                query.Add(new Term("ContentDescription", word));
            }
            //query.Add(new Term("body", "语言")); --可以添加查询条件,两者是add关系.顺序没有关系.
            // query.Add(new Term("body", "大学生"));
            //query.Add(new Term("body", kw));//body中含有kw的文章
            query.Slop = 100;//多个查询条件的词之间的最大距离.在文章中相隔太远 也就无意义.(例如 “大学生”这个查询条件和"简历"这个查询条件之间如果间隔的词太多也就没有意义了。)
            //TopScoreDocCollector是盛放查询结果的容器
            TopScoreDocCollector collector = TopScoreDocCollector.Create(100, true);

            searcher.Search(query, null, collector);                               //根据query查询条件进行查询,查询结果放入collector容器
            ScoreDoc[] docs = collector.TopDocs(0, collector.TotalHits).ScoreDocs; //得到所有查询结果中的文档,GetTotalHits():表示总条数   TopDocs(300, 20);//表示得到300(从300开始),到320(结束)的文档内容.
            List <ViewModelContent> ViewModelList = new List <ViewModelContent>();

            for (int i = 0; i < docs.Length; i++)
            {
                //
                //搜索ScoreDoc[]只能获得文档的id,这样不会把查询结果的Document一次性加载到内存中。降低了内存压力,需要获得文档的详细内容的时候通过searcher.Doc来根据文档id来获得文档的详细内容对象Document.
                int              docId            = docs[i].Doc;         //得到查询结果文档的id(Lucene内部分配的id)
                Document         doc              = searcher.Doc(docId); //找到文档id对应的文档详细信息
                ViewModelContent viewModelContent = new ViewModelContent();
                viewModelContent.Id    = Convert.ToInt32(doc.Get("Id"));
                viewModelContent.Title = doc.Get("Title");
                viewModelContent.ContentDescription = WebCommon.CreateHightLight(Request["txtSearch"], doc.Get("ContentDescription"));
                ViewModelList.Add(viewModelContent);
            }
            SearchDetails searchDetails = new SearchDetails()
            {
                Id             = Guid.NewGuid(),
                KeyWords       = Request["txtSearch"],
                SearchDateTime = DateTime.Now
            };

            SearchDetailsService.AddEntity(searchDetails);
            return(ViewModelList);
        }
Beispiel #5
0
        public async Task <ActionResult> Search([FromServices] IPostService postService, string wd = "", [Range(1, int.MaxValue, ErrorMessage = "页码必须大于0")] int page = 1, [Range(1, 50, ErrorMessage = "页大小必须在0到50之间")] int size = 15)
        {
            wd = ChineseConverter.Convert(wd?.Trim() ?? "", ChineseConversionDirection.TraditionalToSimplified);
            ViewBag.PageSize = size;
            ViewBag.Keyword  = wd;
            string key = "Search:" + ClientIP;

            if (CacheManager.Exists(key) && CacheManager.Get(key) != wd)
            {
                var hotSearches = RedisHelper.Get <List <KeywordsRank> >("SearchRank:Week").Take(10).ToList();
                ViewBag.hotSearches = hotSearches;
                ViewBag.ErrorMsg    = "10秒内只能搜索1次!";
                return(View(new SearchResult <PostDto>()));
            }

            if (!string.IsNullOrWhiteSpace(wd) && !wd.Contains("锟斤拷"))
            {
                new JiebaSegmenter().AddWord(wd);
                if (!HttpContext.Session.TryGetValue("search:" + wd, out _) && !HttpContext.Request.IsRobot())
                {
                    SearchDetailsService.AddEntity(new SearchDetails
                    {
                        Keywords   = wd,
                        SearchTime = DateTime.Now,
                        IP         = ClientIP
                    });
                    await SearchDetailsService.SaveChangesAsync();

                    HttpContext.Session.Set("search:" + wd, wd.ToByteArray());
                }

                var posts = postService.SearchPage(page, size, wd);
                if (posts.Total > 1)
                {
                    CacheManager.AddOrUpdate(key, wd, s => wd);
                    CacheManager.Expire(key, TimeSpan.FromSeconds(10));
                    ViewBag.Ads = AdsService.GetByWeightedPrice(AdvertiseType.PostList);
                }

                ViewBag.hotSearches = new List <KeywordsRank>();
                return(View(posts));
            }

            ViewBag.hotSearches = RedisHelper.Get <List <KeywordsRank> >("SearchRank:Week").Take(10).ToList();
            return(View(new SearchResult <PostDto>()));
        }