This repository that hold the information about the competition about validating and fixing relationship direction in Cypher statements based on the given schema.
- https://github.com/GolvenLD/cypher-direction-competition-entry (Python, RegEx)
- https://github.com/sakusaku-rich/cypher-direction-competition (Python, RegEx)
- https://github.com/grapergrape/cypher-competition (Python)
- https://github.com/mrtnzagustin/cypher-direction-challenge/tree/master (Python, RegEx)
- https://github.com/Frodnar/cypher_direction_validation_entry (Python, RegEx)
- https://github.com/yWorks/cypher-query-validator (TypeScript, Cypher Parser)
- https://github.com/bSharpCyclist/cypher-direction-competition (Python, RegEx)
- https://github.com/sarmbruster/cypher-direction (Python, Cypher Parser)
- https://github.com/nielsdejong/cypher-guard (Python, Cypher Parser)
- https://github.com/filantrop/cypher_fix_relation_directions (Python, Cypher Parser)
- https://github.com/neo4j/jbang-catalog/blob/main/enforce_relationship_directions.java (Java, Cypher Parser)
In the time of LLMs, it is becoming increasingly popular to implement Retrieval-Augmented Generation (RAG) applications, where you feed additional information to an LLM at query time to increase its ability to generate valid, accurate, and up-to-date answers. This new RAG paradigm is bringing in the revolution in data accessibility, as it means that anybody, even non-technical users, can now ask questions about information stored in the database without requiring them to learn a database query language. For example, when you want the users to be able to ask questions about the information stored in Neo4j, a graph database, you need to implement a so-called text2cypher module that takes natural language as input and produces Cypher statements as output. State-of-the-art LLMs are pretty good at generating Cypher statements. However, most of them share a common flaw: they can sometimes mess up the direction of the relationship in the generated Cypher statement, which can cause significant dissatisfaction among users. I believe that the direction of the relationship can be deterministically after the LLM generates a Cypher statement based on the provided schema. Therefore, I am hosting this competition to help us find the best implementation of validating and fixing relationship directions in Cypher statements as accurately and fast as possible.
The idea is to use the winner's code and add it to LLM libraries like LangChain, LlamaIndex, and others. By applying to this competition, you are allowing me, or others, to re-use the provided code in any commercial or non-commercial application with appropriate attribution.
- Python is preferred (although applications in other languages will be accepted)
- No external libraries or tools shall be used. Only standard, bundled libraries apply. The only exceptions are Cypher AST parsers.
The solutions will be judged based on the following three criterias, in the order given:
- Accuracy of results
- Simplicity of the code
- Readability of the code
- Performance of the code
Prizes are the following:
- 1500€
- 750€
- 250€
Sponsored by Neo4j <3
Winners' code will be added to this repository.
Fill out the form with your name and the link to a public GitHub repository with the implementation code: https://forms.gle/4pgArm8S8NqRa6sy8 All applications received until Friday, 17th September 2023, 23.59 CEST time will be valid.
The examples.csv contains the dataset you can use to validate your implementation.
The CSV contains the following three columns:
- statement: Input Cypher statement that is used as the input to your implementation
- schema: Given graph schema that you can use to validate relationship direction and correct them if needed
- correct_query: Expected output
The given schema is given as a list of triples where:
[(Person, WORKS_AT, Organization)]
- first element of the triple specifies the source or start node of a relationship:
Person - second element of the triple specifies the relationship type:
WORKS_AT - third element of the triple specifies the target or end node of a relationship:
Organization
If the given pattern in a Cypher statement doesn't fit the graph schema, simply return an empty string, (there are two examples like this in the test dataset)
Please let me know if you find any bugs in the dataset!
When I was preparing the Cypher examples, I followed these guidelines:
-
If the given pattern in a Cypher statement doesn't fit the graph schema, simply return an empty string
-
If the relationship is between two nodes of the same labels, there is nothing to validate or correct
(:Person)-->(:Person), (:Person)-[:KNOWS]->(:Person)
- If the input query has an undirected relationship in the pattern, we do not correct it.
(:Person)--(:Organization), (:Person)-[:WORKS_AT]-(:Organization)
- If a node label is missing in the defined pattern, we can still validate if it fits the graph schema
(:Person)-[:WORKS_AT]->()
- If the input query doesn't define the relationship type, but at least one node label is given of a pattern, we check if any relationship exists that matches the pattern and correct it if needed
(:Person)-->(), (:Organization)<-[r]-()
- When multiple relationships are given or a negation is used in a pattern, make sure that at least one of relationship types of the possible fits the given schema
(:Person)-[:KNOWS|WORKS_AT]->(:Organization), (:Person)-[:!KNOWS]->(:Organization)
- When variable length patten is used, we do not correct the direction or validate the schema
(:Person)-[:WORKS_AT*]->(:Person), (:Person)-[:WORKS_AT*1..4]->(:Person)
- Node labels or relationship types can optionally be wrapped with backticks
(:`Person`)-[:`WORKS_AT`]->(:Organization)
- Graph schema never contains information about multi-labeled nodes, but the input query can have them. For example, the following input query:
MATCH (a:Person:Actor)-[:ACTED_IN]->(:Movie)
RETURN a, count(*)
Will have an accompanying schema:
[(Person, ACTED_IN, Movie), (Actor, ACTED_IN, Movie)]
Which means that you can use any of the labels of the multi-labeled node to verify relationship directions
All rights reserved to change the conditions of the competition if needed or see fit.