Some cleanup

modules/joex/src/main/resources/reference.conf

@@ -299,9 +299,8 @@ docspell.joex {
       # multiple are given, they are all tried and the "best" is
       # chosen at the end. See
       # https://nlp.stanford.edu/nlp/javadoc/javanlp/edu/stanford/nlp/classify/ColumnDataClassifier.html
-      # for more info about these settings. The settings are almost
-      # identical to them, as they yielded best results with *my*
-      # dataset.
+      # for more info about these settings. The settings here yielded
+      # good results with *my* dataset.
       #
       # Enclose regexps in triple quotes.
       classifiers = [
@@ -312,6 +311,7 @@ docspell.joex {
           "maxNGramLeng" = "4"
           "minNGramLeng" = "1"
           "splitWordShape" = "chris4"
+          "intern" = "true" # makes it slower but saves memory
         }
       ]
     }

modules/joex/src/main/scala/docspell/joex/learn/LearnClassifierTask.scala

@@ -20,6 +20,7 @@ import bitpeace.MimetypeHint
 
 object LearnClassifierTask {
   val noClass = "__NONE__"
+  val pageSep = " --n-- "
 
   type Args = LearnClassifierArgs
 
@@ -80,7 +81,9 @@ object LearnClassifierTask {
     val connStream =
       for {
         item <- QItem.findAllNewesFirst(ctx.args.collective, 10).through(restrictTo(max))
-        tt   <- Stream.eval(QItem.resolveTextAndTag(ctx.args.collective, item, category))
+        tt <- Stream.eval(
+          QItem.resolveTextAndTag(ctx.args.collective, item, category, pageSep)
+        )
       } yield Data(tt.tag.map(_.name).getOrElse(noClass), item.id, tt.text.trim)
     ctx.store.transact(connStream.filter(_.text.nonEmpty))
   }

modules/joex/src/main/scala/docspell/joex/process/TextAnalysis.scala

@@ -11,6 +11,7 @@ import docspell.analysis.nlp.TextClassifier
 import docspell.common._
 import docspell.joex.Config
 import docspell.joex.analysis.RegexNerFile
+import docspell.joex.learn.LearnClassifierTask
 import docspell.joex.process.ItemData.AttachmentDates
 import docspell.joex.scheduler.Context
 import docspell.joex.scheduler.Task
@@ -76,7 +77,7 @@ object TextAnalysis {
     for {
       model <- findActiveModel(ctx, cfg)
       _     <- OptionT.liftF(ctx.logger.info(s"Guessing tag …"))
-      text = metas.flatMap(_.content).mkString("   ------   ")
+      text = metas.flatMap(_.content).mkString(LearnClassifierTask.pageSep)
       modelData =
         ctx.store.bitpeace
           .get(model.id)
@@ -89,8 +90,7 @@ object TextAnalysis {
           .compile
           .drain
           .flatMap(_ => classifier.classify(ctx.logger, ClassifierModel(modelFile), text))
-
-      })
+      }).filter(_ != LearnClassifierTask.noClass)
       _ <- OptionT.liftF(ctx.logger.debug(s"Guessed tag: ${cls}"))
     } yield cls
 

modules/store/src/main/scala/docspell/store/queries/QItem.scala

@@ -634,7 +634,8 @@ object QItem {
   def resolveTextAndTag(
       collective: Ident,
       itemId: Ident,
-      tagCategory: String
+      tagCategory: String,
+      pageSep: String
   ): ConnectionIO[TextAndTag] = {
     val aId    = RAttachment.Columns.id.prefix("a")
     val aItem  = RAttachment.Columns.itemId.prefix("a")
@@ -682,7 +683,7 @@ object QItem {
         s"Got ${texts.size} text and tag entries for item ${itemId.id}"
       )
       tag = texts.headOption.flatMap(_._2)
-      txt = texts.map(_._1).mkString(" --n-- ")
+      txt = texts.map(_._1).mkString(pageSep)
     } yield TextAndTag(itemId, txt, tag)
   }