Skip to content

Introduction to local knowledge graph

Mini lab - Building a local Knowledge Graph RAG with Neo4j, LangChain, and Ollama 🦙

In this lab we will be building a small knowledge graph from Wikipedia, loading it into Neo4j, and querying it with natural language using Langchain library and a local LLM (Ollama).

We will try to replicate the following architecture from the official neo4j blog with a local LLM through ollama app.

Learning goals 🎯

  • use an external api
  • extract entities and relationships from text into a knowledge graph
  • load graph data into Neo4j and inspect it with Cypher
  • play with Cypher and Langchain

👉 You will be uploading your results inside your personnal PUBLIC github repository, as usal this is a personnal project.

You can deliver this lab in a jupyter notebook format (and optionnal python scripts, but you will need to have a notebook at the root of you github repo to demonstrate your solution)

Prerequisites & setup

  • Python 3.10+
  • Neo4j available locally (via Docker)
  • Ollama running locally with some model pulled (e.g., llama3:8b)
  • A FREE Diffbot API key for entity/relationship extraction (NO CB REQUIRED, ⚠️ do not put your cb ⚠️). You can set it as an environment variable DIFFBOT_API_KEY, or edit the notebook cell that initializes DiffbotGraphTransformer.

1) Create and activate a virtual environment : 2) Install the given dependencies : 

pip install -U langchain langchain-experimental langchain-openai langchain-neo4j neo4j wikipedia langchain-community langchain-text-splitters
3) Start Neo4j with the docker-compose file below 
services:
  neo4j:
    image: neo4j:5.20
    environment:
      NEO4J_AUTH: neo4j/password
      NEO4J_PLUGINS: '["apoc","graph-data-science","bloom"]'
      NEO4J_dbms_security_procedures_unrestricted: "apoc.*,gds.*"
      NEO4J_server_config_strict__validation_enabled: "false"
      NEO4J_dbms_default__database: shop
    ports: ["7474:7474","7687:7687"]
    volumes:
      - ./neo4j/data:/data
      - ./neo4j/import:/import

4) Start Ollama and pull a model:

ollama serve
# in another terminal
ollama pull llama3:8b

Walkthrough

  • Ensure core packages are installed inside the notebook environment (be aware of dep conflict)

  • Load a Wikipedia articles with this lanchain module from some names or famous character and extract the documents with Diffbot api

    You can choose any person or topic for example WikipediaLoader(query="Satoshi Nakamoto")

  • Connect to Neo4j and ingest the graph: clean the DB first it's not empty, and add the extracted graph documents from diffbot.

  • Then you can query relationships around Satoshi (or your topic/person)

    Run a Cypher query to inspect relationships from :Person nodes relating to “Satoshi Nakamoto” for example.

  • Do some graph exploration and analysis : list node labels, inspect raw relationships, graph visualization

  • Write graph exploration queries (top connections, interests, employment, etc.)Run several Cypher snippets to explore the graph (e.g., most connected people, interests, employment, locations, competitors, founded-by).

  • Verify that the target entity exists and see how it’s represented (if multiple “Satoshi Nakamoto” nodes exist, how would you choose the "one"?)

  • Write a python reusable function to query all relationships for a person.

  • Create a GraphCypherQAChain with a custom Cypher prompt and ChatOllama model.

    What elements in the custom prompt help the model generate correct Cypher for your schema?

  • Wrap the chain invocation to simplify later usage with a ask_graph(question) helper function.

  • Implement get_entity_subgraph and summarize_entity functions to build structured context and prompt an LLM for summaries.

    What information from the graph would you add/remove to improve the quality and factuality of summaries?

  • Question : Does the LLM hallucinate dates or facts? How would you constrain it to avoid this?

Some tips

  • Dependency warnings: If you see langchain-core incompatibilities, align versions across langchain, langchain-community, langchain-openai, langchain-neo4j, langchain-experimental, and langchain-text-splitters. Pinning versions in a requirements.txt file will help you 🥸
  • Neo4j schema: If queries return empty, verify that labels and properties (:Person, .id, .name) match the ones created by your extractor you can install the vscode plugin
  • Relationship unions: Prefer MATCH (p)-[:EMPLOYEE_OR_MEMBER_OF|:WORK_RELATIONSHIP]->(o)
  • Ollama: ensure ollama serve is running and your model is pulled. If responses are weak, try a larger model or adjust the prompt

Happy coding and remeber : RTFM 🤗