entitykb 21.7.28

Python toolkit for rapidly developing knowledge bases

💡 EntityKB

Python toolkit for the rapid development of custom knowledge bases.

---

EntityKB Documentation: https://www.entitykb.org

EntityKB Code Repository: https://github.com/genomoncology/entitykb

EntityKB Python Package: https://pypi.org/project/entitykb/

---

Overview

EntityKB is a toolkit for rapidly developing knowledge bases (i.e. knowledge graphs) using the Python programming language. It combines an text processing pipeline with a graph data store to support rules/string-based entity extraction and linking.

Potential Use Cases

• Entity Extraction: Pull concepts from unstructured text using using keyword and pattern matching.

• Entity Linking: Map concepts to a knowledge graph for "semantic searching" capabilities for recommendation systems, Q&A, data harmonization, and validated data entry (i.e. drop downs, typeahead).

• Data Set Labeling: Overcome the "cold start" training set problem by using entity extraction capabilities to generate annotations.

• Knowledge Curation/Integration: Iteratively add new data types and sources with simple Python model classes and data loaders.

Capabilities

EntityKB provides a focused set of core capabilities that can be extended and enhanced:

• Graph data store for storing of entities (nodes) and their relationships (edges) as plain-old python objects.

• Terms and Edge index for efficient retrieval and traversal of entities using their name, synonyms, and relationships.

• Processing pipeline that normalizes and tokenizes text and then resolves entities using string, regex, grammar or python based matching.

• Searching with fluent, pythonic traversal query builder for walking and filtering graph nodes and their relationships.

• Importing and exporting of data with CLI tooling and/or Python code.

• Multiple interfaces including embedded Python client, RPC/HTTP servers and CLI.

Priorities

EntityKB's core goal is to enable the rapid, iterative development of custom knowledge graphs using Python. The following quality attributes have been prioritized in furtherance of this goal:

• Evolvability: Add, update and remove entity types and data sources without time-consuming data migrations.

• Configurability: Activate and deactivate out-of-the-box components and custom code by editing a simple JSON file.

• Interoperability: Interact via command-line, RPC, HTTP, or in-memory Python library based on evolving project needs.

• Understandability: Create new entity classes, contextual labels/verbs, and custom resolvers with domain specific language and concepts.

• Portability: Code and data created for EntityKB should be transferable to a new technology stack with minimal effort.

Limitations

EntityKB is deliberately limited in scope to minimize complexity. Below are some choices that users should be aware of upfront:

• Not secure: EntityKB has no authentication or authorization capabilities. RPC and HTTP services should not be exposed to untrusted clients. Instead, proxy EntityKB behind your application's security layer. Also, knowledge bases are stored using pickles which are not secure, only unpickle data from trusted sources.

• Python-based: The most critical data structures are powered by libraries written in the C programming language for high-performance and tight memory management. However, this library is still mostly Python-based and will certainly not meet everyone's runtime needs.

• Not transactional: EntityKB is not designed for ACID-compliant data storage and should never be used as the "system of record". EntityKB can be updated during runtime, but care should be taken to prevent data loss or corruption.

• Not ML based: EntityKB is a software development platform without any out-of-the-box machine learning capabilities. However, it certainly can be used in larger ML-based projects and custom resolvers can be added that use ML models for their entity detection logic.

• Not Resilient: EntityKB graph searching capability has no guards against long-running queries that can impact system responsiveness. Limit end user's from creating open-ended queries to prevent service disruption.

---

Getting Started

Install

$ pip install entitykb

Initialize

EntityKB init creates a KB in the specified "root" directory. The root directory is determined using the following priorities:

1. Command-line argument
2. Environment variable (ENTITYKB_ROOT)
3. Default path (~/.entitykb)

Below are the init and info commands using the default path. Notice the default configuration specifies implementation classes that can be overridden using the config.json file. The index.db contains the graph and terms index data in python pickle format and can be deployed with the config.json to any server using the same version of EntityKB.

$ entitykb init
INFO:     Initialization completed successfully.

$ ls ~/.entitykb/
config.json
edges
edges.dawg
nodes
nodes.dawg

$ cat ~/.entitykb/config.json
{
    "graph": "entitykb.Graph",
    "modules": [],
    "normalizer": "entitykb.LatinLowercaseNormalizer",
    "searcher": "entitykb.DefaultSearcher",
    "terms": "entitykb.TermsIndex",
    "tokenizer": "entitykb.WhitespaceTokenizer",
    "pipelines": {
        "default": {
            "extractor": "entitykb.DefaultExtractor",
            "resolvers": [
                "entitykb.TermResolver"
            ],
            "filterers": []
        }
    }
}

$ entitykb info
+------------------------------------+-----------------------------------+
| config.graph                       |                    entitykb.Graph |
| config.modules                     |                                   |
| config.normalizer                  | entitykb.LatinLowercaseNormalizer |
| config.pipelines.default.extractor |         entitykb.DefaultExtractor |
| config.pipelines.default.filterers |                                   |
| config.pipelines.default.resolvers |             entitykb.TermResolver |
| config.root                        |        /Users/ianmaurer/.entitykb |
| config.searcher                    |          entitykb.DefaultSearcher |
| config.terms                       |               entitykb.TermsIndex |
| config.tokenizer                   |      entitykb.WhitespaceTokenizer |
| entitykb.version                   |                          20.12.14 |
| graph.edges                        |                                 0 |
| graph.nodes                        |                                 0 |
+------------------------------------+-----------------------------------+

Interact

Start a new Knowledge Base and add two entities:

>>> from entitykb import KB, Entity
>>> kb = KB()

>>> nys = Entity(name="New York", label="STATE")
>>> kb.save_node(nys)

>>> nyc = Entity(name="New York City", label="CITY", synonyms=["NYC"])
>>> kb.save_node(nyc)
Entity(key='New York City|CITY', label='CITY', data=None, name='New York City', synonyms=('NYC',))

Edge objects can be created using the Edge class or the right/left-shift methods on nodes. The Verb (or short-hand V) can be used or so can a simple string. The library converts all verbs to uppercase:

>>> from entitykb import Edge, Verb, V 

>>> Edge(start="New York City|CITY", verb=V.IS_IN, end=nys)
Edge(start='New York City|CITY', verb='IS_IN', end='New York|STATE')

>>> nyc >> Verb.IS_IN >> nys
Edge(start='New York City|CITY', verb='IS_IN', end='New York|STATE')

>>> nys << V.is_in << nyc
Edge(start='New York City|CITY', verb='IS_IN', end='New York|STATE')

>>> nys << "is_in" << "New York City|CITY"
Edge(start='New York City|CITY', verb='IS_IN', end='New York|STATE')

Edges still need to be saved to the graph which can be accomplished using the save_edge or connect methods:

>>> edge = nyc >> Verb.IS_IN >> nys
>>> kb.save_edge(edge)

>>> kb.connect(start=nyc, verb="IS_IN", end=nyc)
Edge(start='New York|STATE', verb='IS_IN', end='New York City|CITY')

To index the terms and edges of the Graph store, you must first run the reindex method on the KB:

>>> kb.reindex()

Perform term search using common prefix text:

>>> response = kb.search("New Y")
>>> len(response)
2
>>> response[0]
Entity(key='New York|STATE', label='STATE', data=None, name='New York', synonyms=())
>>> response[1]
Entity(key='New York City|CITY', label='CITY', data=None, name='New York City', synonyms=('NYC',))

Parse text into a document with tokens and spans containing entities:

>>> doc = kb.parse("NYC is another name for New York City")
>>> len(doc.tokens)
8
>>> doc.spans
(NYC, New York City)
>>> doc.entities
(Entity(key='New York City|CITY', label='CITY', data=None, name='New York City', synonyms=('NYC',)),
Entity(key='New York City|CITY', label='CITY', data=None, name='New York City', synonyms=('NYC',)))

The dump command generates JSONL:

$ entitykb dump
{"kind": "node", "payload": {"key": "New York|STATE", "label": "STATE", "data": null, "name": "New York", "synonyms": []}}
{"kind": "node", "payload": {"key": "New York City|CITY", "label": "CITY", "data": null, "name": "New York City", "synonyms": ["NYC"]}}
{"kind": "edge", "payload": {"__root__": ["New York City|CITY", "IS_IN", "New York|STATE"]}}

Dump and load the data for safe transfer to a different version of EntityKB:

$ entitykb dump /tmp/out.jsonl
$ wc -l /tmp/out.jsonl
2
$ entitykb clear
Are you sure you want to clear: /Users/ianmaurer/.entitykb/index.db? [y/N]: y
INFO:     Clear completed successfully.

$ # do version switch here.

$ $ entitykb load /tmp/out.jsonl
Loaded 2 in 0.02s [/tmp/out.jsonl, jsonl]
Reindexed in 0.01s

---

Background

History

EntityKB was developed by GenomOncology and is the foundation of the clinical, molecular and genomic knowledge base that power GenomOncology's igniteIQ data extraction platform and clinical decision support API suite. EntityKB was released as an open source library in November 2020 for the benefit of GenomOncology's clients and the greater open source community.

Maintainer

The initial version of EntityKB was designed and implemented by Ian Maurer who is the Chief Technology Officer (CTO) for GenomOncology. Ian has over 20 years of industry experience and is the architect of GenomOncology's igniteIQ data extract platform and API Suite that powers GenomOncology's Precision Oncology Platform.

Ian can be contacted via Twitter, LinkedIn, or email (ian -at- genomoncology.com).

Related Projects

EntityKB was inspired by and is powered by several other projects in the open source community. Below are the most salient examples:

• DAWG is used for indexing edge relationships (triples), terms, and labels. It supports "completion" of terms based on a prefix.

• DiskCache is used for storing nodes and edges to disk as python objects.

• Pydantic for model annotations, schema definition and FastAPI documentation.

• Typer powers EntityKB's Command Line Interface (CLI) tool.

• FastAPI powers EntityKB's HTTP Application Programming Interface (API).

• uvicorn and Starlette for power running FastAPI.

• MkDocs, Termynal.js, and Material for MkDocs for making the documentation look great.

• Lark for powering the date resolver's grammar.

• FlashText for inspiring parts of EntityKB's design and approach.

License

This project is copyrighted by GenomOncology and licensed under the terms of the MIT license.

---

Status

EntityKB should be considered beta software. Some caveats:

• Expect backwards incompatible changes that will break your Knowledge Base. When that happens, run the following steps: 1. Using the old version of EntityKB, run the dump CLI command to generate a JSONL file of your edges and nodes. 2. Switch to the new version of EntityKB, and run the load CLI command to import the edges and nodes from JSONL.

• Monitor changes in the release notes.

• To minimize frustration, please pin the version of the software in your requirements.txt or equivalent file.

---

Contributing

• Submit bugs and enhancement suggestions via GitHub issues.

• Contributions welcome, please see Development for setting up a working dev environment.

• Our goal is to keep EntityKB's code footprint as small as possible but contrib modules will be more readily accepted.

• Separate packages don't require a pull request, but please follow the naming pattern entitykb-<name> to aid in discoverability on pypi.