Exploration of using LLMs to create knowledge graphs
- go 1.24+
- python 3.10+
- uv
- docker
- make
- httpie
Spin up the application with make up
This will start:
- The API server on port
8000 - The Neo4j database on port
7687(bolt protocol) - The Neo4j browser on port
7474
Build a knowledge graph with the Python CLI by following the instructions in the scripts/README.md file.
This will populate the Neo4j database with the initial subjects and then progressively populate the database with more related subjects. An OpenAI API key is required to run the script.
Cypher queries can be run in the Neo4j browser here.
Example query showing all subjects and their related subjects:
MATCH (s1:Subject {kgVersion: "1"})
OPTIONAL MATCH (s1)-[r:RELATED_TO]-(s2:Subject)
RETURN s1, r, s2Spin down the application with make down
The most general subjects tend to be the ones most referenced as a related subject, but this is not always the case.
MATCH (s1:Subject {kgVersion: "2"})
RETURN s1.name, COUNT { (s1)-[:RELATED_TO]->(s2) } AS num_related_subjects
ORDER BY num_related_subjects DESC
LIMIT 15Querying for subjects with the most related subject references:
MATCH (s1:Subject {kgVersion: "2"})
RETURN s1.name, COUNT { (s1)-[:RELATED_TO]->(s2) } AS num_related_subjects
ORDER BY num_related_subjects ASC
LIMIT 15Querying for subjects with the least related subject references:
// Find tightly connected groups of 3
MATCH (s1:Subject {kgVersion: "2"})-[:RELATED_TO]-(shared:Subject),
(s2:Subject {kgVersion: "2"})-[:RELATED_TO]-(shared),
(s3:Subject {kgVersion: "2"})-[:RELATED_TO]-(shared)
WHERE s1 <> s2 AND s2 <> s3 AND s1 <> s3
AND elementId(s1) < elementId(s2) AND elementId(s2) < elementId(s3)
WITH s1, s2, s3, count(DISTINCT shared) AS shared_connections
WHERE shared_connections >= 3
RETURN s1.name, s2.name, s3.name, shared_connections
ORDER BY shared_connections DESC
LIMIT 20
This query can be expanded as follows:
// Find tightly connected groups of 5 subjects with shared neighbors
MATCH (s1:Subject {kgVersion: "2"})-[:RELATED_TO]-(shared:Subject),
(s2:Subject {kgVersion: "2"})-[:RELATED_TO]-(shared),
(s3:Subject {kgVersion: "2"})-[:RELATED_TO]-(shared),
(s4:Subject {kgVersion: "2"})-[:RELATED_TO]-(shared),
(s5:Subject {kgVersion: "2"})-[:RELATED_TO]-(shared)
WHERE s1 <> s2 AND s1 <> s3 AND s1 <> s4 AND s1 <> s5
AND s2 <> s3 AND s2 <> s4 AND s2 <> s5
AND s3 <> s4 AND s3 <> s5
AND s4 <> s5
AND elementId(s1) < elementId(s2)
AND elementId(s2) < elementId(s3)
AND elementId(s3) < elementId(s4)
AND elementId(s4) < elementId(s5)
WITH s1, s2, s3, s4, s5, count(DISTINCT shared) AS shared_connections
WHERE shared_connections >= 3
RETURN s1.name, s2.name, s3.name, s4.name, s5.name, shared_connections
ORDER BY shared_connections DESC
LIMIT 20See notes here for potential improvements.
- BUGS:
- Subject special characters are not escaped before insertion into the database
- Update population script to expose CLI options for:
- KG version
- Number of related subjects to fetch per subject
- Max Subject count
- Create a few more knowledge graph versions
- See if the clustering is similar for each KG version
- Iterate on LLM prompt
- Have it only get progressively more specific as it goes deeper
- Have it return some parent and some children