diff --git a/src/rcldb/rcldb.cpp b/src/rcldb/rcldb.cpp
index 806e3f93..e1131808 100644
--- a/src/rcldb/rcldb.cpp
+++ b/src/rcldb/rcldb.cpp
@@ -559,13 +559,18 @@ int Db::Native::getPageNumberForPosition(const vector<int>& pbreaks,
     return it - pbreaks.begin() + 1;
 }
 
+// Note: we're passed a Xapian::Document* because Xapian can't do
+// reference-counting properly. We take ownership and need to delete
+// it before returning.
 bool Db::Native::addOrUpdateWrite(const string& udi, const string& uniterm, 
-				  Xapian::Document& newdocument, size_t textlen)
+				  Xapian::Document *newdocument_ptr, 
+                                  size_t textlen)
 {
 #ifdef IDX_THREADS
     Chrono chron;
     PTMutexLocker lock(m_mutex);
 #endif
+    RefCntr<Xapian::Document> doc_cleaner(newdocument_ptr);
 
     // Check file system full every mbyte of indexed text. It's a bit wasteful
     // to do this after having prepared the document, but it needs to be in
@@ -590,8 +595,11 @@ bool Db::Native::addOrUpdateWrite(const string& udi, const string& uniterm,
     // Add db entry or update existing entry:
     try {
 	Xapian::docid did = 
-	    xwdb.replace_document(uniterm, newdocument);
+	    xwdb.replace_document(uniterm, *newdocument_ptr);
 	if (did < m_rcldb->updated.size()) {
+            // This is necessary because only the file-level docs are tested
+            // by needUpdate(), so the subdocs existence flags are only set
+            // here.
 	    m_rcldb->updated[did] = true;
 	    LOGINFO(("Db::add: docid %d updated [%s]\n", did, fnc));
 	} else {
@@ -604,7 +612,7 @@ bool Db::Native::addOrUpdateWrite(const string& udi, const string& uniterm,
 	ermsg.erase();
 	// FIXME: is this ever actually needed?
 	try {
-	    xwdb.add_document(newdocument);
+	    xwdb.add_document(*newdocument_ptr);
 	    LOGDEB(("Db::add: %s added (failed re-seek for duplicate)\n", 
 		    fnc));
 	} XCATCHERROR(ermsg);
@@ -1245,7 +1253,12 @@ bool Db::addOrUpdate(const string &udi, const string &parent_udi, Doc &doc)
     if (m_ndb == 0)
 	return false;
 
-    Xapian::Document newdocument;
+    // This document is potentially going to be passed to the index
+    // update thread. The reference counters are not mt-safe, so we
+    // need to do this through a pointer. The reference is just there
+    // to avoid changing too much code (the previous version passed a copy).
+    Xapian::Document *newdocument_ptr = new Xapian::Document;
+    Xapian::Document &newdocument(*newdocument_ptr);
     
     // The term processing pipeline:
     TermProcIdx tpidx;
@@ -1270,8 +1283,10 @@ bool Db::addOrUpdate(const string &udi, const string &parent_udi, Doc &doc)
 	// first. This is so different from the normal processing that
 	// it uses a fully separate code path (with some duplication
 	// unfortunately)
-	if (!m_ndb->docToXdocXattrOnly(&splitter, udi, doc, newdocument))
+	if (!m_ndb->docToXdocXattrOnly(&splitter, udi, doc, newdocument)) {
+            delete newdocument_ptr;
 	    return false;
+        }
     } else {
 
 	// If the ipath is like a path, index the last element. This is
@@ -1365,6 +1380,7 @@ bool Db::addOrUpdate(const string &udi, const string &parent_udi, Doc &doc)
 		     "count %d avglen %.4f sigma %.4f url [%s] ipath [%s] text %s\n",
 		     v.count, v.avglen, v.sigma, doc.url.c_str(), 
 		     doc.ipath.c_str(), doc.text.c_str()));
+            delete newdocument_ptr;
 	    return true;
 	}
 #endif
@@ -1551,9 +1567,10 @@ bool Db::addOrUpdate(const string &udi, const string &parent_udi, Doc &doc)
 #ifdef IDX_THREADS
     if (m_ndb->m_havewriteq) {
 	DbUpdTask *tp = new DbUpdTask(DbUpdTask::AddOrUpdate, udi, uniterm, 
-				      newdocument, doc.text.length());
+				      newdocument_ptr, doc.text.length());
 	if (!m_ndb->m_wqueue.put(tp)) {
 	    LOGERR(("Db::addOrUpdate:Cant queue task\n"));
+            delete newdocument_ptr;
 	    return false;
 	} else {
 	    return true;
@@ -1561,7 +1578,7 @@ bool Db::addOrUpdate(const string &udi, const string &parent_udi, Doc &doc)
     }
 #endif
 
-    return m_ndb->addOrUpdateWrite(udi, uniterm, newdocument, 
+    return m_ndb->addOrUpdateWrite(udi, uniterm, newdocument_ptr,
 				   doc.text.length());
 }
 
@@ -1959,7 +1976,7 @@ bool Db::purgeFile(const string &udi, bool *existed)
 #ifdef IDX_THREADS
     if (m_ndb->m_havewriteq) {
 	DbUpdTask *tp = new DbUpdTask(DbUpdTask::Delete, udi, uniterm, 
-				      Xapian::Document(), (size_t)-1);
+				      0, (size_t)-1);
 	if (!m_ndb->m_wqueue.put(tp)) {
 	    LOGERR(("Db::purgeFile:Cant queue task\n"));
 	    return false;
@@ -1986,7 +2003,7 @@ bool Db::purgeOrphans(const string &udi)
 #ifdef IDX_THREADS
     if (m_ndb->m_havewriteq) {
 	DbUpdTask *tp = new DbUpdTask(DbUpdTask::PurgeOrphans, udi, uniterm, 
-				      Xapian::Document(), (size_t)-1);
+				      0, (size_t)-1);
 	if (!m_ndb->m_wqueue.put(tp)) {
 	    LOGERR(("Db::purgeFile:Cant queue task\n"));
 	    return false;
diff --git a/src/rcldb/rcldb_p.h b/src/rcldb/rcldb_p.h
index fda1b018..b247c21a 100644
--- a/src/rcldb/rcldb_p.h
+++ b/src/rcldb/rcldb_p.h
@@ -51,14 +51,14 @@ public:
     // passed both just to avoid recomputing uniterm which is
     // available on the caller site.
     DbUpdTask(Op _op, const string& ud, const string& un, 
-	      const Xapian::Document &d, size_t tl)
+	      Xapian::Document *d, size_t tl)
 	: op(_op), udi(ud), uniterm(un), doc(d), txtlen(tl)
     {}
     // Udi and uniterm equivalently designate the doc
     Op op;
     string udi;
     string uniterm;
-    Xapian::Document doc;
+    Xapian::Document *doc;
     // txtlen is used to update the flush interval. It's -1 for a
     // purge because we actually don't know it, and the code fakes a
     // text length based on the term count.
@@ -101,7 +101,7 @@ class Db::Native {
 
     // Final steps of doc update, part which need to be single-threaded
     bool addOrUpdateWrite(const string& udi, const string& uniterm, 
-			  Xapian::Document& doc, size_t txtlen);
+			  Xapian::Document *doc, size_t txtlen);
 
     /** Delete all documents which are contained in the input document, 
      * which must be a file-level one.