Docria provides a hypergraph document model implementation with a focus on NLP (Natural Language Processing) applications.

Docria provides:

  • In-memory object representations in Python and Java

  • Binary serialization format based on MessagePack

  • File formats optimized for storing and accessing millions of documents locally and in a cluster context


To install the PyPI version:

pip install docria

To install the development version:

git clone
cd docria
pip install -e .

The first steps

from docria.model import Document, DataTypes as T
import regex as re

# Stupid tokenizer
tokenizer = re.compile(r"[a-zA-Z]+|[0-9]+|[^\s]")
starts_with_uppercase = re.compile(r"[A-Z].*")

doc = Document()

# Create a new text context called 'main' with the text 'This code was written in Lund, Sweden.'
doc.maintext = "This code was written in Lund, Sweden."
#               01234567890123456789012345678901234567
#               0         1         2         3
main_text = doc.maintext

# Create a new layer with fields: id, text and head.
# Fields:
#   id is an int32
#   uppercase is a boolean indicating if the token is uppercase
#   text is a textspan from context 'main'
tokens = doc.add_layer("token", id=T.int32(), uppercase=T.bool(), text=T.span())

# Adding nodes: Solution 1
i = 0
for m in tokenizer.finditer(str(main_text)):
    token_node = tokens.add(id=i, text=main_text[m.start():m.end()])

    # Check if it is uppercase
    token_node["uppercase"] = starts_with_uppercase.fullmatch(m[0]) is not None
    i += 1

# Reading nodes
for tok in tokens:

# Filtering, only uppercase tokens
for tok in tokens[tokens["uppercase"] == True]:


The document model consists of the following concepts:

  • Document: The overall container for everything (all nodes, layers, texts must be contained within)

  • Document properties: a single dictionary per document to store metadata in.

  • Text: The basic text representation, a wrapped string to track spans.

  • Text Spans: Subsequence of a string, can always be converted into a hard string by using str(span)

  • Node Spans: Start and stop node in a layer which will produce a sequence of nodes.

  • Layer: Collection of nodes

  • Layer Schema: Definition of field names and types when document is serialized

  • Node: Single node with zero or more fields with values

  • Node fields: Key, value pairs.

from docria.model import Document

doc = Document()
doc.maintext # alias to doc.text["main"] with special support for
             # creating a main text via doc.maintext = "string"

doc.props  # Document metadata dictionary
doc.layers # Layer dictionary, layer name to node layer collection
doc.layer  # Alias to above
doc.texts  # Text dictionary.
doc.text   # Alias to above


Reading document collections

from import MsgpackDocumentReader
from docria.codec import MsgpackDocument

with MsgpackDocumentReader(open("path_to_your_docria_file.docria", "rb")) as reader:
   for rawdoc in reader:
      # rawdoc is of type MsgpackDocument
      doc = rawdoc.document() #  type: docria.Document

      # Print the schema

      for token in doc["token"]:
         # ... do something with the data contained within.

# You can use MsgpackDocumentReader as a normal instance
# and manually use .close() when done or on the GC to eat it up.

The principle is mostly the same with :class:~`` with the exception it expects a filepath, not a filelike object.

Writing document collections

from import MsgpackDocumentReader
from docria.codec import MsgpackDocument

with MsgpackDocumentWriter(open("path_to_your_docria_file.docria", "wb")) as writer:
   # using the previous doc in "The first steps"

# Rewriting or filtering
with MsgpackDocumentWriter(open("path_to_your_output_docria_file.docria", "wb")) as writer:
   with MsgpackDocumentReader(open("path_to_your_input_docria_file.docria", "rb")) as reader:
      for rawdoc in reader:
         writer.write(rawdoc)  # this is decompression and memory copy of the raw data

The principle is mostly the same with :class:~`` with the exception it expects a filepath, not a filelike object.

Reading and writing documents to bytes

from docria.codec import MsgpackCodec, MsgpackDocument

binarydata = bytes()  # from any location
binarydata = io.BytesIO()  # or

# To decode a msgpack document into a document
msgdoc = MsgpackDocument(binarydata)
doc = msgdoc.document()  # type: docria.model.Document

# To encode a document into a msgpack document
msgdoc = MsgpackDocument(doc)
binarydata = msgdoc.binary()  # type: bytes

# Access data without a full deserialization
rawdoc = MsgpackDocument(binarydata)  # Document metadata as dictionary

# Document texts, dictionary name to list of strings
# (each segment which potentially has annotation) which can be joined to get the full text.

schema = rawdoc.schema() # advanced access to the contents of this document, lists layers and fields.

doc = rawdoc.document() # full document deserialization

Layer and field query

from docria import Document, DataTypes as T, NodeSpan, NodeList

doc = Document()
doc.maintext = "Lund is a city in Sweden."
#               0123456789012345678901234
#               0         1         2

# Only ordered layers exist in docria, this means all nodes are added sequentially.
# T.span() is equivalent to T.span("main") which referes to the main text
token_layer = doc.add_layer("token", part_of_speech=T.string(), text=T.span(), head=T.noderef("token"))

# Annotation output by CoreNLP 3.9.2 and Basic dependencies
# We set node references later.
first = token_layer.add(part_of_speech="NNP", text=doc.maintext[0:4])
token_layer.add(part_of_speech="VBZ", text=doc.maintext[5:7])
token_layer.add(part_of_speech="DT", text=doc.maintext[8:9])
token_layer.add(part_of_speech="NN", text=doc.maintext[10:14])
token_layer.add(part_of_speech="IN", text=doc.maintext[15:17])
token_layer.add(part_of_speech="NNP", text=doc.maintext[18:24])
last = token_layer.add(part_of_speech=".", text=doc.maintext[24:])

# Create a node span and convert into a list
sent_tokens = NodeSpan(first, last).to_list()

# When setting heads, no validation takes place.
sent_tokens[0]["head"] = token_layer[3] # head = city
sent_tokens[1]["head"] = token_layer[3] # head = city
sent_tokens[2]["head"] = token_layer[3] # head = city
sent_tokens[4]["head"] = token_layer[5] # head = Sweden
sent_tokens[5]["head"] = token_layer[3] # head = city
sent_tokens[6]["head"] = token_layer[3] # head = city

sent_tokens.validate() # We can manually initiate validate for these nodes to fail faster.

# This first query finds all roots by checking if the head is None, and finally picks the first one.
first_root = token_layer[token_layer["head"].is_none()].first()

# This second query finds all nodes with the head equal to first_root
tokens_with_head_first_root = token_layer[token_layer["head"] == first_root]

# Then we print tokens in layer order from matching token to including root token
for tok in tokens_with_head_first_root:
    # iter_span is invariant to order, it will always produce low id to high id.

Change presentation settings

The settings used for pretty printing is controlled by the global variable docria.printout.options which is a docria.printout.PrintOptions.

By convention pretty printing will output [layer name]#[internal id] where the internal id can be used to get the node. However, this id is only guaranteed to be static if the layer is not changed, if changed it is invalid.

For references in general use the Node object.

API Reference / Modules


Docria document model ( primary module )


Functions for various processing purposes


Codecs, encoding/decoding documents to/from binary or text representations


I/O module, read/write collections of documents


Presentation module, utilities for formatting document objects.

Indices and tables