diff --git a/doc/src/file-icatdata.rst b/doc/src/file-icatdata.rst
index 2dbb00c..06d8f70 100644
--- a/doc/src/file-icatdata.rst
+++ b/doc/src/file-icatdata.rst
@@ -29,18 +29,19 @@ Data files are partitioned in chunks.  This is done to avoid having
 the whole file, e.g. the complete inventory of the ICAT, at once in
 memory.  The problem is that objects contain references to other
 objects (e.g. Datafiles refer to Datasets, the latter refer to
-Investigations, and so forth).  We keep an index of the objects in
-order to resolve these references.  But there is a memory versus time
-tradeoff: we cannot keep all the objects in the index, that would
-again mean the complete inventory of the ICAT.  And we can't know
-beforehand which object is going to be referenced later on, so we
-don't know which one to keep and which one to discard from the index.
-Fortunately we can query objects we discarded once back from the ICAT
-server.  But this is expensive.  So the strategy is as follows: keep
-all objects from the current chunk in the index and discard the
-complete index each time a chunk has been processed.  This will work
-fine if objects are mostly referencing other objects from the same
-chunk and only a few references go across chunk boundaries.
+Investigations, and so forth).  We keep an index of the objects as
+cache in order to resolve these references.  But there is a memory
+versus time tradeoff: we cannot keep all the objects in the index,
+that would again mean the complete inventory of the ICAT.  And we
+can't know beforehand which object is going to be referenced later on,
+so we don't know which one to keep and which one to discard from the
+index.  Fortunately we can query objects that we discarded once back
+from the ICAT server.  But this is expensive.  So the strategy is as
+follows: keep all objects from the current chunk in the index and
+discard the complete index each time a chunk has been
+processed. [#dc]_ This will work fine if objects are mostly
+referencing other objects from the same chunk and only a few
+references go across chunk boundaries.
 
 Therefore, we want these chunks to be small enough to fit into memory,
 but at the same time large enough to keep as many relations between
@@ -48,18 +49,24 @@ objects as possible local in a chunk.  It is in the responsibility of
 the writer of the data file to create the chunks in this manner.
 
 The objects that get written to the data file and how this file is
-organized is controlled by lists of ICAT search expressions, see
-:meth:`icat.dumpfile.DumpFileWriter.writeobjs`.  There is some degree
-of flexibility: an object may include related objects in an
-one-to-many relation, just by including them in the search expression.
-In this case, these related objects should not have a search
-expression on their own again.  For instance, the search expression
-for Grouping may include UserGroup.  The UserGroups will then be
-embedded in their respective grouping in the data file.  There should
-not be a search expression for UserGroup then.
+organized is controlled by lists of ICAT search expressions or entity
+objects, see :meth:`icat.dumpfile.DumpFileWriter.writeobjs`.  There is
+some degree of flexibility: an object may include related objects in
+an one-to-many relation.  In this case, these related objects should
+not be added on their own again.  For instance, you may write User,
+Grouping, and UserGroup as separate objects into the file.  In this
+case, the UserGroup entries must properly reference related User and
+Grouping.  Alternatively you may include the UserGroups in the
+corresponding Grouping objects.  In this case, you must not add the
+UserGroups again on their own.
 
 Objects related in a many-to-one relation must always be included in
 the search expression.  This is also true if the object is
 indirectly related to one of the included objects.  In this case,
 only a reference to the related object will be included in the data
-file.  The related object must have its own list entry.
+file.  The related object must have its own entry.
+
+
+.. [#dc] There is one exception: DataCollections don't have a
+         uniqueness constraint and can't reliably be searched by
+         attributes.  They are always kept in the index.