A Graph-based Retrieval-Augmented Generation system that ingests documents, builds a Neo4j knowledge graph with Cypher, and uses Gemma on the Groq platform for fast, accurate, relationship-aware question answering.

Traditional RAG retrieves chunks of unstructured text using search techniques such as dense vector similarity, sparse/lexical search (e.g., keyword or BM25), or hybrid search that combines both approaches. Graph RAG goes further — it stores data as entities (nodes) and relationships (edges) in a graph database, enabling multi-hop reasoning(the ability to connect and traverse multiple linked facts to answer complex queries) and delivering explainable answers.

This project:

1. Ingests documents.
2. Extracts entities and relationships.
3. Stores them in Neo4j.
4. Uses LangChain’s GraphCypherQAChain to query the graph.
5. Passes relevant context to Gemma (via Groq) for final answer generation.

• Graph Database (Neo4j): Stores and queries data as nodes & edges for connected insights.
• Knowledge Graph: Structured network of facts linking entities and relationships.
• RAG: Retrieval-Augmented Generation — retrieve external data, feed to an LLM.
• Graph RAG: RAG enhanced with graph queries for deeper, relationship-aware reasoning.

flowchart LR
    A[Document Ingestion] --> B[Entity & Relationship Extraction]
    B --> C[Cypher Query Generation]
    C --> D[Neo4j Knowledge Graph]
    E[User Query] --> F[GraphCypherQAChain]
    D --> F
    F --> G[Gemma LLM via Groq]
    G --> H[Context-Aware Answer]

pip install --upgrade langchain langchain-community langchain-groq neo4j

export NEO4J_URI="bolt://<host>:7687"
export NEO4J_USERNAME="neo4j"
export NEO4J_PASSWORD="<password>"
export GROQ_API_KEY="<groq-api-key>"

from langchain_community.graphs import Neo4jGraph
from langchain_groq import ChatGroq
from langchain.chains import GraphCypherQAChain
import os

graph = Neo4jGraph(url=os.environ["NEO4J_URI"],
                   username=os.environ["NEO4J_USERNAME"],
                   password=os.environ["NEO4J_PASSWORD"])
graph.refresh_schema()

llm = ChatGroq(groq_api_key=os.environ["GROQ_API_KEY"], model_name="Gemma2-9b-It")

chain = GraphCypherQAChain.from_llm(llm=llm, graph=graph, verbose=True, allow_dangerous_requests=True)

result = chain.invoke({"query": "Who was the director of the movie GoldenEye"})
print(result)

LOAD CSV WITH HEADERS FROM 'https://raw.githubusercontent.com/.../movies_small.csv' AS row
MERGE (m:Movie {id: row.movieId})
SET m.title = row.title, m.released = date(row.released), m.imdbRating = toFloat(row.imdbRating)
FOREACH (actor IN split(row.actors, '|') |
  MERGE (p:Person {name: trim(actor)}) MERGE (p)-[:ACTED_IN]->(m))
FOREACH (director IN split(row.directors, '|') |
  MERGE (p:Person {name: trim(director)}) MERGE (p)-[:DIRECTED]->(m))
FOREACH (genre IN split(row.genres, '|') |
  MERGE (g:Genre {name: trim(genre)}) MERGE (m)-[:IN_GENRE]->(g));

Database visualisation in Graph : (you can see here https://console-preview.neo4j.io/tools/query )

Database visualisation in Table :

The screenshot above shows the reasoning steps and final answer generated by the Gemma model after retrieving relevant nodes and relationships from Neo4j.

✅ Multi-hop reasoning over connected facts ✅ More accurate, explainable answers ✅ Works well in finance, healthcare, research, legal domains

• Neo4j — Graph database
• Cypher — Graph query language
• Gemma — Large Language Model
• Groq — High-speed inference
• LangChain — Orchestration

• Use environment variables or secret managers for credentials.
• allow_dangerous_requests=True allows generated Cypher execution — validate queries in production.
• Enhance ingestion with NLP-based entity/relation extraction for better graph quality.