cve-kgrag-db / code /src /constructors /neo4j_graph_service.py
DuyTa's picture
Add code/: full CVE-KGRAG project source snapshot
27f6252 verified
Raw
History Blame Contribute Delete
27.8 kB
#!/usr/bin/env python3
"""
Neo4j Graph Service — replaces NetworkX graph builder.
Responsibilities
----------------
1. bulk_load() — load all KG JSON files into Neo4j in batched transactions.
2. similar_cves() — find CVEs related by shared CWE / product / vendor.
3. subgraph_for_cves() — expand a set of CVE IDs by N hops for RAG context.
4. (optional) centrality_scores() — PageRank via GDS for ranked context.
Run as CLI to do the bulk load:
python -m src.constructors.neo4j_graph_service --bulk-load
python -m src.constructors.neo4j_graph_service --similar CVE-2021-44228
python -m src.constructors.neo4j_graph_service --stats
"""
from __future__ import annotations
import argparse
import json
import logging
import sys
from pathlib import Path
from typing import Any, Dict, List, Optional, Tuple
from neo4j import GraphDatabase, Session
project_root = Path(__file__).parent.parent.parent
sys.path.insert(0, str(project_root))
from config import Config
from src.generators.rag_config import NEO4J_URI, NEO4J_USER, NEO4J_PASSWORD
logging.basicConfig(level=logging.INFO, format="%(asctime)s - %(levelname)s - %(message)s")
logger = logging.getLogger(__name__)
_BATCH = 10_000 # rows per Neo4j transaction
class Neo4jGraphService:
"""
Bulk-load KG JSON → Neo4j and expose graph query helpers.
Connects on construction; call .close() when done.
"""
def __init__(
self,
uri: str = NEO4J_URI,
user: str = NEO4J_USER,
pwd: str = NEO4J_PASSWORD,
) -> None:
self.driver = GraphDatabase.driver(uri, auth=(user, pwd))
self.config = Config()
self.kg_dir = self.config.knowledge_base_dir / "knowledge_graph"
def close(self) -> None:
self.driver.close()
def __enter__(self) -> "Neo4jGraphService":
return self
def __exit__(self, *_: Any) -> None:
self.close()
# ── schema ────────────────────────────────────────────────────────────────
def create_constraints(self) -> None:
"""Unique constraints guarantee idempotent MERGE."""
constraints = [
("CVE", "id"),
("Product", "name"),
("Vendor", "name"),
("CWE", "id"),
("CAPEC", "id"),
("Technique", "id"),
("Tactic", "id"),
]
with self.driver.session() as s:
for label, prop in constraints:
try:
s.run(
f"CREATE CONSTRAINT IF NOT EXISTS "
f"FOR (n:{label}) REQUIRE n.{prop} IS UNIQUE"
)
except Exception as e:
logger.warning(f"Constraint {label}.{prop}: {e}")
logger.info("Constraints ensured.")
# ── bulk loader ───────────────────────────────────────────────────────────
def bulk_load(self) -> None:
"""Load all JSON KG files into Neo4j. Idempotent (uses MERGE)."""
logger.info("Starting Neo4j bulk load …")
self.create_constraints()
self._load_cve_nodes()
self._load_product_nodes()
self._load_vendor_nodes()
self._load_cwe_nodes()
self._load_capec_nodes()
self._load_technique_nodes()
self._load_tactic_nodes()
self._load_rels("cve_cwe_relationships.json", self._upsert_cve_cwe)
self._load_rels("cve_product_relationships.json", self._upsert_cve_product)
self._load_rels("cve_capec_relationships.json", self._upsert_cve_capec)
self._load_rels("cve_mitre_technique_relationships.json",self._upsert_cve_technique)
self._load_rels("cve_mitre_tactic_relationships.json", self._upsert_cve_tactic)
self._load_rels("product_vendor_relationships.json", self._upsert_product_vendor)
logger.info("Bulk load complete.")
self._log_counts()
# ── node loaders ──────────────────────────────────────────────────────────
def _load_cve_nodes(self) -> None:
data = self._read_json("cves_nodes.json")
rows = list(data.values()) if isinstance(data, dict) else data
logger.info(f"Loading {len(rows):,} CVE nodes …")
cypher = """
UNWIND $rows AS r
MERGE (c:CVE {id: r.id})
SET c.title = r.title,
c.description = r.description,
c.source = r.source,
c.published_date = r.published_date,
c.last_modified = r.last_modified_date,
c.cvss_v3_score = r.cvss_v3_score,
c.cvss_v3_severity = r.cvss_v3_severity,
c.cvss_v3_vector = r.cvss_v3_vector,
c.attack_vector = r.attack_vector,
c.attack_complexity = r.attack_complexity,
c.privileges_required = r.privileges_required,
c.user_interaction = r.user_interaction,
c.scope = r.scope,
c.conf_impact = r.conf_impact,
c.integ_impact = r.integ_impact,
c.avail_impact = r.avail_impact,
c.cvss_v2_score = r.cvss_v2_score,
c.cvss_v2_vector = r.cvss_v2_vector,
c.is_in_kev = r.is_in_kev,
c.exploitdb_count = r.exploitdb_count,
c.tags = r.tags
"""
def _norm(r: Dict[str, Any]) -> Dict[str, Any]:
cvss3 = r.get("cvss_v3") or {}
cvss2 = r.get("cvss_v2") or {}
# Use prose description (first line of content if available)
raw_desc = r.get("content", r.get("description", "")) or ""
prose = raw_desc.split("\n")[0].strip()
# Build a real title from the first sentence of the description
title = prose[:120] if prose else r.get("id", "")
return {
"id": r.get("id", ""),
"title": title,
"description": prose,
"source": r.get("source", ""),
"published_date": r.get("published_date", ""),
"last_modified_date":r.get("last_modified_date", ""),
"cvss_v3_score": cvss3.get("base_score"),
"cvss_v3_severity": cvss3.get("base_severity", ""),
"cvss_v3_vector": cvss3.get("vector_string", ""),
"attack_vector": cvss3.get("attack_vector", ""),
"attack_complexity": cvss3.get("attack_complexity", ""),
"privileges_required": cvss3.get("privileges_required", ""),
"user_interaction": cvss3.get("user_interaction", ""),
"scope": cvss3.get("scope", ""),
"conf_impact": cvss3.get("confidentiality_impact", ""),
"integ_impact": cvss3.get("integrity_impact", ""),
"avail_impact": cvss3.get("availability_impact", ""),
"cvss_v2_score": cvss2.get("base_score"),
"cvss_v2_vector": cvss2.get("vector_string", ""),
"is_in_kev": bool(r.get("is_in_kev", False)),
"exploitdb_count": r.get("exploitdb_correlations_count", 0) or 0,
"tags": r.get("tags") or [],
}
self._batch_write(cypher, [_norm(r) for r in rows])
def _load_product_nodes(self) -> None:
data = self._read_json("products_nodes.json")
rows = list(data.values()) if isinstance(data, dict) else data
logger.info(f"Loading {len(rows):,} Product nodes …")
cypher = """
UNWIND $rows AS r
MERGE (p:Product {name: r.name})
SET p.vendor = r.vendor,
p.display_name = r.display_name,
p.category = r.category,
p.criticality_score = r.criticality_score
"""
self._batch_write(cypher, rows)
def _load_vendor_nodes(self) -> None:
data = self._read_json("vendors_nodes.json")
rows = list(data.values()) if isinstance(data, dict) else data
logger.info(f"Loading {len(rows):,} Vendor nodes …")
cypher = """
UNWIND $rows AS r
MERGE (v:Vendor {name: r.name})
"""
# vendors_nodes may store as {vendor_name: {name:..., ...}}
norm_rows = []
for r in rows:
norm_rows.append({"name": r.get("name", r.get("vendor", ""))})
self._batch_write(cypher, norm_rows)
def _load_cwe_nodes(self) -> None:
data = self._read_json("cwes_nodes.json")
rows = list(data.values()) if isinstance(data, dict) else data
logger.info(f"Loading {len(rows):,} CWE nodes …")
# Enrich with real names + descriptions from cwe_chunks.json if available
cwe_enrichment: Dict[str, Dict[str, str]] = {}
cwe_chunks_path = (
self.config.knowledge_base_dir / "rag_exports" / "cwe_chunks.json"
)
if cwe_chunks_path.exists():
try:
with open(cwe_chunks_path, encoding="utf-8") as f:
cwe_chunks = json.load(f)
for c in cwe_chunks:
p = c.get("payload", {})
cid = p.get("cwe_id", "")
if cid:
cwe_enrichment[cid] = {
"name": p.get("name", ""),
"description": c.get("text", "")[:500],
}
logger.info(f" CWE enrichment loaded: {len(cwe_enrichment)} entries")
except Exception as e:
logger.warning(f" Could not load CWE enrichment: {e}")
cypher = """
UNWIND $rows AS r
MERGE (w:CWE {id: r.id})
SET w.name = r.name,
w.description = r.description,
w.vuln_count = r.vulnerability_count
"""
def _norm(r: Dict[str, Any]) -> Dict[str, Any]:
cid = r.get("id", "")
enriched = cwe_enrichment.get(cid, {})
raw_name = r.get("name", "")
# Strip verbose prefix "Common Weakness Enumeration CWE-XXX" → real name
clean_name = (
enriched.get("name")
or raw_name.replace("Common Weakness Enumeration ", "").strip()
or cid
)
return {
"id": cid,
"name": clean_name,
"description": enriched.get("description", ""),
"vulnerability_count": r.get("vulnerability_count", 0),
}
self._batch_write(cypher, [_norm(r) for r in rows])
def _load_capec_nodes(self) -> None:
data = self._read_json("capecs_nodes.json")
rows = list(data.values()) if isinstance(data, dict) else data
logger.info(f"Loading {len(rows):,} CAPEC nodes …")
cypher = """
UNWIND $rows AS r
MERGE (a:CAPEC {id: r.id})
SET a.name = r.name
"""
self._batch_write(cypher, rows)
def _load_technique_nodes(self) -> None:
data = self._read_json("mitre_techniques_nodes.json")
rows = list(data.values()) if isinstance(data, dict) else data
logger.info(f"Loading {len(rows):,} Technique nodes …")
# Load real names + descriptions from CTI technique JSON files
cti_dir = self.config.cti_docs_dir / "attack_techniques"
technique_enrichment: Dict[str, Dict[str, str]] = {}
if cti_dir.exists():
for jf in cti_dir.glob("*.json"):
try:
with open(jf, encoding="utf-8") as f:
doc = json.load(f)
tid = doc.get("id", "")
if tid:
technique_enrichment[tid] = {
"name": doc.get("title", ""),
"description": (doc.get("content") or "")[:400],
"tactics": ", ".join(doc.get("tactics", [])),
}
except Exception:
pass
logger.info(f" Technique enrichment loaded: {len(technique_enrichment)} entries")
cypher = """
UNWIND $rows AS r
MERGE (t:Technique {id: r.id})
SET t.name = r.name,
t.description = r.description,
t.tactics = r.tactics,
t.vuln_count = r.vulnerability_count
"""
def _norm(r: Dict[str, Any]) -> Dict[str, Any]:
tid = r.get("id", "")
enriched = technique_enrichment.get(tid, {})
raw_name = r.get("name", r.get("title", ""))
clean_name = (
enriched.get("name")
or raw_name.replace("MITRE ATT&CK Technique ", "").strip()
or tid
)
return {
"id": tid,
"name": clean_name,
"description": enriched.get("description", ""),
"tactics": enriched.get("tactics", ""),
"vulnerability_count": r.get("vulnerability_count", 0),
}
self._batch_write(cypher, [_norm(r) for r in rows])
def _load_tactic_nodes(self) -> None:
data = self._read_json("mitre_tactics_nodes.json")
rows = list(data.values()) if isinstance(data, dict) else data
logger.info(f"Loading {len(rows):,} Tactic nodes …")
cypher = """
UNWIND $rows AS r
MERGE (t:Tactic {id: r.id})
SET t.name = r.name,
t.vuln_count = r.vulnerability_count
"""
def _norm(r: Dict[str, Any]) -> Dict[str, Any]:
raw_name = r.get("name", r.get("title", ""))
# Strip verbose prefix "MITRE ATT&CK Tactic TXXX" → real name
clean_name = raw_name.replace("MITRE ATT&CK Tactic ", "").strip() or r.get("id", "")
return {
"id": r.get("id", ""),
"name": clean_name,
"vulnerability_count": r.get("vulnerability_count", 0),
}
self._batch_write(cypher, [_norm(r) for r in rows])
# ── relationship loaders ──────────────────────────────────────────────────
def _load_rels(self, filename: str, handler) -> None:
path = self.kg_dir / filename
if not path.exists():
logger.warning(f"Relationship file not found: {filename}")
return
with open(path, encoding="utf-8") as f:
rows = json.load(f)
logger.info(f"Loading {len(rows):,} {filename} …")
handler(rows)
def _upsert_cve_cwe(self, rows: List[Dict[str, Any]]) -> None:
cypher = """
UNWIND $rows AS r
MATCH (c:CVE {id: r.cve_id}), (w:CWE {id: r.cwe_id})
MERGE (c)-[:HAS_WEAKNESS]->(w)
"""
self._batch_write(cypher, rows)
def _upsert_cve_product(self, rows: List[Dict[str, Any]]) -> None:
cypher = """
UNWIND $rows AS r
MATCH (c:CVE {id: r.cve_id}), (p:Product {name: r.product_name})
MERGE (c)-[:AFFECTS]->(p)
"""
self._batch_write(cypher, rows)
def _upsert_cve_capec(self, rows: List[Dict[str, Any]]) -> None:
cypher = """
UNWIND $rows AS r
MATCH (c:CVE {id: r.cve_id}), (a:CAPEC {id: r.capec_id})
MERGE (c)-[:USES_PATTERN]->(a)
"""
norm = [{"cve_id": r.get("cve_id", ""), "capec_id": r.get("capec_id", "")} for r in rows]
self._batch_write(cypher, norm)
def _upsert_cve_technique(self, rows: List[Dict[str, Any]]) -> None:
cypher = """
UNWIND $rows AS r
MATCH (c:CVE {id: r.cve_id}), (t:Technique {id: r.technique_id})
MERGE (c)-[:IMPLEMENTS_TECHNIQUE]->(t)
"""
norm = [{"cve_id": r.get("cve_id", ""), "technique_id": r.get("technique_id", "")} for r in rows]
self._batch_write(cypher, norm)
def _upsert_cve_tactic(self, rows: List[Dict[str, Any]]) -> None:
cypher = """
UNWIND $rows AS r
MATCH (c:CVE {id: r.cve_id}), (t:Tactic {id: r.tactic_id})
MERGE (c)-[:MAPPED_TO_TACTIC]->(t)
"""
norm = [{"cve_id": r.get("cve_id", ""), "tactic_id": r.get("tactic_id", "")} for r in rows]
self._batch_write(cypher, norm)
def _upsert_product_vendor(self, rows: List[Dict[str, Any]]) -> None:
cypher = """
UNWIND $rows AS r
MATCH (p:Product {name: r.product_name}), (v:Vendor {name: r.vendor_name})
MERGE (p)-[:MANUFACTURED_BY]->(v)
"""
norm = [{"product_name": r.get("product_name", r.get("product", "")),
"vendor_name": r.get("vendor_name", r.get("vendor", ""))} for r in rows]
self._batch_write(cypher, norm)
# ── query API ─────────────────────────────────────────────────────────────
def similar_cves(self, cve_id: str, k: int = 10) -> List[Dict[str, Any]]:
"""
Find CVEs related to cve_id by shared CWE, product, or tactic.
Returns list of {cve_id, shared_type, score}.
"""
cypher = """
MATCH (src:CVE {id: $cve_id})
CALL {
WITH src
MATCH (src)-[:HAS_WEAKNESS]->(w:CWE)<-[:HAS_WEAKNESS]-(other:CVE)
WHERE other.id <> $cve_id
RETURN other.id AS cve_id, 'cwe' AS shared_type, count(w) AS cnt
UNION
WITH src
MATCH (src)-[:AFFECTS]->(p:Product)<-[:AFFECTS]-(other:CVE)
WHERE other.id <> $cve_id
RETURN other.id AS cve_id, 'product' AS shared_type, count(p) AS cnt
UNION
WITH src
MATCH (src)-[:MAPPED_TO_TACTIC]->(t:Tactic)<-[:MAPPED_TO_TACTIC]-(other:CVE)
WHERE other.id <> $cve_id
RETURN other.id AS cve_id, 'tactic' AS shared_type, count(t) AS cnt
}
RETURN cve_id, shared_type, sum(cnt) AS score
ORDER BY score DESC
LIMIT $k
"""
with self.driver.session() as s:
result = s.run(cypher, cve_id=cve_id, k=k)
return [{"cve_id": r["cve_id"], "shared_type": r["shared_type"], "score": r["score"]}
for r in result]
def subgraph_for_cves(self, cve_ids: List[str], hops: int = 1) -> Dict[str, Any]:
"""Backward-compat wrapper — seed subgraph from CVE IDs."""
return self.subgraph_from_nodes(
[{"label": "CVE", "id": cid} for cid in cve_ids], hops=hops
)
# ── label-agnostic KG API ─────────────────────────────────────────────────
_LABEL_KEY: Dict[str, str] = {
"CVE": "id", "CWE": "id", "CAPEC": "id",
"Technique": "id", "Tactic": "id",
"Product": "name", "Vendor": "name",
}
_LABEL_TO_CVE_REL: Dict[str, Optional[str]] = {
"CWE": "HAS_WEAKNESS",
"CAPEC": "USES_PATTERN",
"Technique": "IMPLEMENTS_TECHNIQUE",
"Tactic": "MAPPED_TO_TACTIC",
"Product": "AFFECTS",
"Vendor": None, # handled via Product hop
}
def lookup_by_label(self, label: str, query: str, k: int = 5) -> List[Dict[str, Any]]:
"""
Fuzzy + exact match a node by label.
Returns [{id, name, props}]. Label must be in _LABEL_KEY whitelist.
"""
if label not in self._LABEL_KEY:
logger.warning("lookup_by_label: unknown label %r", label)
return []
key = self._LABEL_KEY[label]
cypher = f"""
MATCH (n:{label})
WHERE toLower(n.{key}) = toLower($q)
OR toLower(n.{key}) CONTAINS toLower($q)
RETURN coalesce(n.id, n.name) AS id,
coalesce(n.name, n.id) AS name,
properties(n) AS props
LIMIT $k
"""
with self.driver.session() as s:
result = s.run(cypher, q=query, k=k)
return [{"id": r["id"], "name": r["name"], "props": dict(r["props"])} for r in result]
def cves_for_node(self, label: str, node_id: str, k: int = 10) -> List[Dict[str, Any]]:
"""
Given a non-CVE seed (Product, CWE, Technique, Tactic, CAPEC, Vendor),
return connected CVEs ordered by CVSS score desc.
"""
if label not in self._LABEL_TO_CVE_REL:
return []
key = self._LABEL_KEY.get(label, "id")
rel = self._LABEL_TO_CVE_REL[label]
if label == "Vendor":
cypher = f"""
MATCH (v:Vendor)-[:MANUFACTURED_BY]-(p:Product)-[:AFFECTS]-(c:CVE)
WHERE v.name = $nid
RETURN DISTINCT c.id AS id, c.cvss_v3_score AS cvss, c.description AS desc
ORDER BY cvss DESC LIMIT $k
"""
elif rel:
cypher = f"""
MATCH (n:{label})-[:{rel}]-(c:CVE)
WHERE n.{key} = $nid
RETURN c.id AS id, c.cvss_v3_score AS cvss, c.description AS desc
ORDER BY cvss DESC LIMIT $k
"""
else:
return []
with self.driver.session() as s:
return [
{"id": r["id"], "cvss": r["cvss"], "description": r["desc"]}
for r in s.run(cypher, nid=node_id, k=k)
]
def subgraph_from_nodes(self, seeds: List[Dict[str, str]], hops: int = 1) -> Dict[str, Any]:
"""
General subgraph expansion from heterogeneous seeds [{label, id}, ...].
Replaces the CVE-only subgraph_for_cves. Expands each seed label
separately and merges results. Label must be in _LABEL_KEY whitelist.
"""
hops = max(1, int(hops))
by_label: Dict[str, List[str]] = {}
for s in seeds:
lbl = s.get("label", "")
if lbl in self._LABEL_KEY:
by_label.setdefault(lbl, []).append(s["id"])
all_nodes: List[Dict] = []
all_edges: List[Dict] = []
with self.driver.session() as session:
for label, ids in by_label.items():
key = self._LABEL_KEY[label]
cypher = f"""
MATCH path = (seed:{label})-[*1..{hops}]-(n)
WHERE seed.{key} IN $ids
UNWIND nodes(path) AS node
UNWIND relationships(path) AS rel
RETURN DISTINCT
labels(node)[0] AS node_label,
coalesce(node.id, node.name) AS node_id,
node.name AS node_name,
type(rel) AS rel_type,
coalesce(startNode(rel).id, startNode(rel).name) AS rel_start,
coalesce(endNode(rel).id, endNode(rel).name) AS rel_end
LIMIT 500
"""
for r in session.run(cypher, ids=ids):
all_nodes.append({
"label": r["node_label"],
"id": r["node_id"],
"name": r["node_name"],
})
all_edges.append({
"type": r["rel_type"],
"start": r["rel_start"],
"end": r["rel_end"],
})
# Dedup
seen_n: set = set()
seen_e: set = set()
nodes = [n for n in all_nodes if (n["label"], n["id"]) not in seen_n
and not seen_n.add((n["label"], n["id"]))] # type: ignore[func-returns-value]
edges = [e for e in all_edges if (e["type"], e["start"], e["end"]) not in seen_e
and not seen_e.add((e["type"], e["start"], e["end"]))] # type: ignore[func-returns-value]
return {"nodes": nodes, "edges": edges}
# ── stats ─────────────────────────────────────────────────────────────────
def get_stats(self) -> Dict[str, int]:
labels = ["CVE", "Product", "Vendor", "CWE", "CAPEC", "Technique", "Tactic"]
stats: Dict[str, int] = {}
with self.driver.session() as s:
for label in labels:
result = s.run(f"MATCH (n:{label}) RETURN count(n) AS cnt")
stats[label] = result.single()["cnt"]
return stats
def _log_counts(self) -> None:
stats = self.get_stats()
for label, cnt in stats.items():
logger.info(f" {label}: {cnt:,}")
# ── internal ──────────────────────────────────────────────────────────────
def _read_json(self, filename: str) -> Any:
path = self.kg_dir / filename
if not path.exists():
logger.warning(f"File not found: {path}")
return {}
logger.info(f" Reading {filename} ({path.stat().st_size / 1e6:.0f} MB) …")
with open(path, encoding="utf-8") as f:
return json.load(f)
def _batch_write(self, cypher: str, rows: List[Dict[str, Any]]) -> None:
total = len(rows)
with self.driver.session() as s:
for start in range(0, total, _BATCH):
batch = rows[start : start + _BATCH]
s.run(cypher, rows=batch)
if (start + _BATCH) % (10 * _BATCH) == 0:
logger.info(f" … {start + _BATCH:,} / {total:,}")
# ─────────────────────────────────────────────────────────────────────────────
# CLI
# ─────────────────────────────────────────────────────────────────────────────
if __name__ == "__main__":
parser = argparse.ArgumentParser(description="Neo4j Graph Service")
parser.add_argument("--bulk-load", action="store_true", help="Load KG JSON → Neo4j")
parser.add_argument("--similar", type=str, help="Find CVEs similar to given CVE-ID")
parser.add_argument("--k", type=int, default=10)
parser.add_argument("--stats", action="store_true")
args = parser.parse_args()
with Neo4jGraphService() as svc:
if args.bulk_load:
svc.bulk_load()
elif args.similar:
results = svc.similar_cves(args.similar, k=args.k)
print(f"\nCVEs similar to {args.similar}:")
for r in results:
print(f" {r['cve_id']:25s} ({r['shared_type']:10s}) score={r['score']}")
elif args.stats:
stats = svc.get_stats()
print("\nNeo4j node counts:")
for label, cnt in stats.items():
print(f" {label:12s}: {cnt:,}")
else:
parser.print_help()