GraphRAG + Blockchain Provenance on AWS: Relationship-Aware and Tamper-Evident QA

Scenario

A legal and compliance platform has good vector RAG recall but weak multi-hop reasoning. Teams need answerability across entities, obligations, jurisdictions, and time, plus evidence integrity.

Why combine GraphRAG and blockchain-style provenance

GraphRAG improves relationship reasoning. Provenance anchoring improves trust and auditability. Combined, they support both “what the model found” and “why to trust that source path.”

Architecture

Trade-offs

• Higher retrieval quality for relationship-heavy queries.
• Increased ingestion complexity.
• Need robust schema governance for graph evolution.

Step-by-step tutorial

1) Provision graph + vector foundations

aws neptune create-db-cluster \
  --db-cluster-identifier graphrag-cluster \
  --engine neptune \
  --db-subnet-group-name my-neptune-subnets \
  --vpc-security-group-ids sg-0123456789abcdef0

aws opensearchserverless create-collection \
  --name graphrag-vectors \
  --type VECTORSEARCH

aws neptune create-db-cluster `
  --db-cluster-identifier graphrag-cluster `
  --engine neptune `
  --db-subnet-group-name my-neptune-subnets `
  --vpc-security-group-ids sg-0123456789abcdef0

aws opensearchserverless create-collection `
  --name graphrag-vectors `
  --type VECTORSEARCH

2) Extract entities and relations

import re


def extract_entities(text: str):
    orgs = re.findall(r"\b[A-Z][A-Za-z0-9& ]+(?:LLC|Inc|Ltd|Corp|Bank)\b", text)
    return sorted(set(o.strip() for o in orgs))


def build_relations(entities: list[str]):
    relations = []
    for i in range(len(entities)-1):
        relations.append((entities[i], "RELATED_TO", entities[i+1]))
    return relations

3) Load graph edges

from gremlin_python.driver import client

g = client.Client("wss://<neptune-endpoint>:8182/gremlin", "g")


def upsert_edge(src, rel, dst):
    q = f"""
    g.V().has('Entity','name','{src}').fold().coalesce(unfold(), addV('Entity').property('name','{src}')).as('a')
     .V().has('Entity','name','{dst}').fold().coalesce(unfold(), addV('Entity').property('name','{dst}')).as('b')
     .coalesce(__.select('a').outE('{rel}').where(inV().as('b')), __.addE('{rel}').from('a').to('b'))
    """
    g.submit(q).all().result()

4) Query orchestration with graph expansion

from fastapi import FastAPI

app = FastAPI()


def vector_candidates(question: str):
    return ["chunk_12", "chunk_45", "chunk_90"]


def graph_expand(seed_entities: list[str]):
    return ["clause_17", "jurisdiction_uk", "obligation_renewal"]


@app.post("/ask")
def ask(payload: dict):
    q = payload["question"]
    chunks = vector_candidates(q)
    related = graph_expand(["Acme LLC"])
    return {"chunks": chunks, "graph_context": related}

5) Add provenance receipts

import hashlib


def receipt(doc_id: str, chunk_text: str, graph_path: str) -> dict:
    digest = hashlib.sha256((doc_id + chunk_text + graph_path).encode("utf-8")).hexdigest()
    return {"doc_id": doc_id, "graph_path": graph_path, "digest": digest}

6) Security and governance

• isolate graph write and read permissions
• enforce schema migration approvals
• sign ingestion artifacts
• maintain document lineage metadata

Monitoring and quality

Track:

• graph expansion depth and latency
• citation completeness
• grounded answer rate
• graph schema drift events

Cost optimization

• cap graph traversal depth for default path
• warm caches for frequent entities
• run heavy enrichment offline

Pricing reminder: verify current pricing for Neptune, OpenSearch Serverless, Bedrock, and S3.

Production checklist

• Graph schema and ontology reviewed by domain experts
• Ingestion is idempotent and replay-safe
• Retrieval and proof receipts logged for each answer
• Security tests include poisoning and relationship tampering
• Cost dashboard includes graph and vector components

References

• https://docs.aws.amazon.com/architecture-diagrams/latest/knowledge-graphs-and-graphrag-with-neo4j/knowledge-graphs-and-graphrag-with-neo4j.html
• https://docs.aws.amazon.com/prescriptive-guidance/latest/retrieval-augmented-generation-options/choosing-option.html
• https://docs.aws.amazon.com/prescriptive-guidance/latest/choosing-an-aws-vector-database-for-rag-use-cases/introduction.html