Build vs buy: graph tools for credit teams

The first four parts of this series argued that private credit questions are often graph questions, built a small cross-lender graph on public data, and walked through sector use cases. This last part covers tooling. We don't rank anything here. Products change quickly (one on this list was archived in the past year), and the right choice depends on your data volume, cloud, team skills and compliance constraints. Check every capability against current vendor documentation before you commit.

Start smaller than you think

Illustrative structure

• Lender / BDC
• Fund / vehicle
• Borrower
• Facility / loan
• Sponsor
• Agent bank
• Amendment
• Filing

The graph in part 2 has about 2,000 nodes and 2,600 edges and ran in seconds in networkx, an in-memory Python library. Most single-firm credit graphs (a few thousand borrowers, tens of thousands of facilities and holdings) are small by graph-database standards. Before you procure anything, ask:

• Is the hard part the database or the entity resolution? For most credit teams it's the second.
• Do several people or applications need to query it concurrently, with access control? If not, a notebook plus files may be enough for a pilot.
• Will LLM agents query it? Then you want a query language the agent can generate reliably, plus a schema you can describe in a prompt.

Graph databases

Neo4j

• What it is: A property-graph database queried with Cypher, offered self-managed and as a managed cloud service (neo4j.com).
• Relevant to credit teams: Large ecosystem and a lot of GraphRAG tooling, including the neo4j-graphrag Python package for building graphs from text and combining vector and graph retrieval. The GraphRAG paper (Edge et al., 2024) lists Neo4j among the libraries with GraphRAG-style integrations.
• Ask about: Which features sit in community vs. enterprise editions (clustering, role-based access), licensing terms, and data residency for the managed service.

Kuzu (archived) and its forks

• What it was: An embedded, in-process property-graph database with Cypher, MIT-licensed. You could run it inside a Python process with no server.
• Status: The GitHub repository was archived by its owner on October 10, 2025 and is read-only (GitHub; The Register). Existing releases still work. Community forks include Kineviz's bighorn and LadybugDB.
• Ask about: If you're drawn to the embedded model, look at who maintains the fork, how often it releases, and how it handles security fixes. Build an exit path to another Cypher engine.

Memgraph

• What it is: An in-memory graph database that uses Cypher, with community and enterprise offerings (docs).
• Relevant to credit teams: Cypher compatibility makes it easier to move queries between engines. The in-memory design suits streaming or frequently updated graphs.
• Ask about: Memory sizing for your graph, persistence and backup, and which features are enterprise-only.

TigerGraph

• What it is: A distributed graph analytics platform with its own query language, GSQL, available self-managed and in the cloud (tigergraph.com).
• Relevant to credit teams: Built for large graphs and deep multi-hop analytics. Its financial-services positioning has focused on use cases like fraud and AML, which share the entity-resolution and network-analysis patterns described here.
• Ask about: GSQL learning curve versus Cypher-based tools (and whether your agents can generate it reliably), plus deployment and pricing model.

Amazon Neptune

• What it is: AWS's managed graph database service, supporting property-graph queries (Gremlin, openCypher) and RDF/SPARQL (aws.amazon.com/neptune).
• Relevant to credit teams: If your data platform is already on AWS, it keeps the graph inside your existing security, networking and IAM controls.
• Ask about: Which query languages and analytics features suit your workload, cost at your query volume, and how it fits with your vector store if you're doing GraphRAG.

Other routes

• Stay in Python: networkx for prototypes and batch analytics, with results written back to your warehouse.
• Open-source pipelines: Microsoft GraphRAG handles LLM-based graph extraction and community summaries and can feed a graph store. Expect to tune it for credit documents (part 3).

Entity resolution: data and services

The database is the easy part. Matching "Zendesk Inc" to "ZenDesk" to the right legal entity and family is where credit graphs succeed or fail. Options, from open to commercial:

• LEI / GLEIF: Free global legal entity identifiers with parent-relationship data. Coverage is strongest for entities that trade or report in regulated markets and patchy for private middle-market borrowers.
• PermID: LSEG's open permanent identifiers for organizations and instruments, with entity-matching tools. Check coverage of private companies for your book.
• Kensho Link (S&P Global): A commercial entity-matching service that maps your company names and attributes to S&P Capital IQ identifiers. Worth testing on a sample of your hardest middle-market names.
• Orbis (Bureau van Dijk, a Moody's company): Commercial private-company data with corporate ownership structures. Useful for parent and subsidiary roll-ups and guarantor families.
• Credit-specific data and news vendors: Several vendors cover private credit deals, documents, LMEs and lender groups, often with their own entity mappings. Capabilities and licensing differ widely. Ask each one directly about entity IDs, coverage of your borrowers, how mappings are maintained, and whether you can export the crosswalk.

Whatever you buy, you'll still need an internal crosswalk: source name → canonical entity → family, with reviewer, date and evidence. Vendor IDs become attributes on your canonical entity, not a replacement for it, because no single vendor covers every private borrower and every data-room spelling.

Build vs buy: questions to settle

1. What's the first question you're answering? Sponsor exposure, CLO overlap, secondaries overlap or document analysis need different data and different tools.
2. Where does the system of record stay? Usually the warehouse or portfolio system. The graph is a derived view, refreshed on a schedule.
3. Who owns entity resolution? Name a person or team. Agents can propose matches, but people approve them.
4. Security and confidentiality. Data-room and fund-finance investor data often come with use restrictions. Model access control and data lineage from day one.
5. Agent integration. Can your LLM agents generate correct queries for the tool's language? Test with your schema. Log every generated query and its result for audit.
6. Exit cost. Prefer open query languages (Cypher/openCypher, SPARQL) and keep the canonical data in your own tables, so switching engines means a reload, not a rewrite. The Kuzu archive is a reminder that this risk is real.
7. Total cost. License or cloud spend is often smaller than the data engineering and review time. Budget for the crosswalk.

A sensible path

• Weeks 1–4: Prototype in Python on one use case, with a reviewed crosswalk for your top few hundred borrowers.
• Months 2–3: If several users or applications need it, load into a graph database your platform team can support, and connect an agent for read-only querying with citations.
• After that: Add vendor IDs where they improve match rates, add document-derived edges (amendments, baskets), and only then consider heavier analytics.

That closes the series. Start with part 1 if you landed here first.

Sources

1. neo4j.com ↗ https://neo4j.com/
2. neo4j.com ↗ https://neo4j.com/docs/neo4j-graphrag-python/current/
3. github.com ↗ https://github.com/kuzudb/kuzu
4. www.theregister.com ↗ https://www.theregister.com/software/2025/10/14/kuzudb-graph-database-abandoned-community-mulls-options/1142229
5. github.com ↗ https://github.com/Kineviz/bighorn
6. blog.ladybugdb.com ↗ https://blog.ladybugdb.com/post/ladybug-spreading-its-wings/
7. memgraph.com ↗ https://memgraph.com/docs
8. www.tigergraph.com ↗ https://www.tigergraph.com/
9. aws.amazon.com ↗ https://aws.amazon.com/neptune/
10. networkx.org ↗ https://networkx.org/
11. www.gleif.org ↗ https://www.gleif.org/en/about-lei/introducing-the-legal-entity-identifier-lei
12. permid.org ↗ https://permid.org/
13. kensho.com ↗ https://kensho.com/link
14. www.bvdinfo.com ↗ https://www.bvdinfo.com/en-gb/our-products/data/international/orbis
15. github.com ↗ https://github.com/microsoft/graphrag
16. arxiv.org ↗ https://arxiv.org/abs/2404.16130

Educational content, not investment advice. Figures derived from SEC filings may contain errors: verify against the filing. Found an error? See the corrections log.