Seven workloads that stop being an integration project

What people run on it.

Most data work is not one question. It is a graph question, then a table question, then a search, then a vector lookup, on the same facts. Today each of those is a separate store, a separate copy and a separate licence. Here they are one.

1. RAG pipelines

Retrieval that keeps the relationships

  • You embed your documents, and you also keep what they are attached to: the contract they belong to, the person who signed it, the ticket that cites them.
  • Retrieval starts as a vector lookup and continues as a walk: the neighbours of the hit, and the neighbours of those.
  • The chunks, the vectors and the graph are the same store, so nothing is copied between a vector database and a graph database to answer one question.
  • Agent memory belongs here too: one shared memory layer that outlives a process. Durability across machines comes with clustering.
Question VECTOR lookup the walk expands Answer + sources
A question enters, a vector lookup finds the first hits, the walk expands to what they are connected to, and the answer leaves with its sources attached.

2. Legal

A matter is a graph, not a folder

  • Matters, parties, contracts, clauses, obligations and dates are all connected, and the useful questions follow the connections: which obligations fall on us if this clause is triggered.
  • Full text search, the document itself and the relationships between documents are one lookup.
  • Everything stays on hardware you control, which is usually the condition of the work.
  • History is kept rather than overwritten, so you can ask what a matter looked like on a date.
CLAUSE triggered Parties Obligations contract to readcontract to readcontract to read
A clause in one contract, traced through the parties and obligations it touches, ending in the documents a lawyer has to read.

3. Banking

Accounts, counterparties and the paths between them

  • Customers, accounts, transfers and counterparties form a graph, and the questions are paths: how money moved from here to there, in how many steps.
  • The same store holds the transaction tables, the customer records and the documents, so a path question does not become an export.
  • Long histories stay queryable instead of being aged out into a second system.
  • Runs inside the bank's own perimeter, with no call out to anyone.
no direct link A B transfertransfertransfertransfer four steps, through three intermediaries
Two accounts with no direct link, and the chain of transfers that connects them through intermediaries.

4. Investments

One store for holdings, prices and the paper behind them

  • Positions, instruments, issuers and ownership chains are relationships: who ultimately owns what, and through which vehicles.
  • Price and time-series data sit beside that, so exposure questions do not span two systems.
  • Research notes, filings and memos are searchable in the same place as the numbers they discuss.
  • Yesterday's view is still there, because history is kept rather than replaced.
HOLDING vehicle vehicle ultimate owner pricesfilingsmemos
A holding traced up its ownership chain to the ultimate owner, with prices and documents hanging off the same nodes.

5. Trading

Ticks and relationships in one place

  • Market data arrives continuously and is kept as a stream you can replay, not only as a table you summarise.
  • Instruments, venues and books are related, so a question about exposure is a walk, not a join across exports.
  • Analytics run over columns of numbers in the same store the stream is written to.
  • What the system saw at a point in time is recoverable, which is what audit and post-mortem actually need.
ticks arrive continuously written once time seriescolumnsinstruments replay a windowaggregatewalk exposure read three ways, from one copy
A stream of ticks landing once, read afterwards as a time series, as columns, and as the instruments they belong to.

6. Utilities

The network, its telemetry and its history

  • The physical network is a graph: assets, feeders, substations, meters, and what depends on what.
  • Telemetry arrives as a continuous stream from those assets and stays attached to the asset that produced it.
  • An outage question is a walk: what is downstream of this failure, and who is affected.
  • Records go back years without moving old data into a second system.
FAILURE customerscustomerscustomers upstream, unaffected everything downstream, and who it reaches
A failure at one asset, with the downstream assets and customers it affects highlighted along the network.

7. Tech

Your own product's data, without the integration project

  • Application data, documents, search, embeddings, events and metrics land in one store with one interface.
  • One thing to run, back up and reason about, instead of a graph database beside a document store beside a search cluster beside a vector index.
  • It installs on your own servers with one command, and the first node is free forever, so evaluating it costs a machine and an afternoon.
  • Clients for the languages you already use.
Your app SEMURG one store one thing to run and back up graph databasedocument storesearch clustervector indexmetrics store replaced
One application and one store, with the stores it replaces greyed out beside it.

Every one of these is the same store answering different questions.

Install a node See the measurements