Architecture
The architecture behind the workspace.
The application people use sits on a connected record. That record supports working memory, approved organisational knowledge and the AI services that need the right context.
Four layers, one record
01
The application
Where people organise, investigate, draft, compare, apply methods and record decisions.
02
The connected record
Evidence, statements, requirements and decisions, and the relationships between them.
03
Memory and knowledge
Working memory for continuity; approved methods and reasoning for later work.
04
Model and integration services
AI and connected systems that receive the context appropriate to a task and a user's access.
SurrealDB as the data platform
Pivot's connected knowledge architecture runs on SurrealDB. Pivot needs to preserve relationships as much as records: which evidence a statement came from, who made it, which requirement a decision addresses and which statements it considered. SurrealDB holds documents, graph relationships and structured queries in one platform, so those connections can be stored and queried together. Professional judgement, permissions, review and the behaviour of the application are Pivot's implementation; the database stores and retrieves the record and does not train models.
The record model
The unified ontology combines definitions for legal work with PAO's treatment of attributed statements. It defines records for participants, evidence, statements, requirements and decisions, and specifies how those records relate. Sector profiles extend it with the terms and outputs each kind of work needs. For example, an expert's valuation remains attributed to that expert when a solicitor selects it for use in a schedule.
Working memory and the record
In Pivot's agent architecture, Agent Memory retains working context such as completed checks and outstanding investigations. The execution design records tool calls, actions and required approvals. Memory is working context; the record holds the evidence and professional decisions against which that context is checked.
Using knowledge to evaluate and improve AI
Context and retrieval
During a task, Pivot can supply the model with relevant evidence and approved guidance from the record, within the user's access. This changes what the model is given to work with, not the model itself.
Evaluation
Curated, expert-reviewed examples test how well a model or agent meets the organisation's standards before it is relied on.
Supported training
Approved examples are prepared as datasets for fine-tuning, with permissions and quality checks applied, for models whose providers support tuning.
Application integration
Pivot's integration layer defines how another application queries the record and requests changes, with permissions and review requirements on every operation. Talk to us about the systems you want to connect.
Questions
Working memory can retain context between sessions, such as checks already performed and work still to do. It is not evidence. Evidence and professional decisions belong in the record, where their sources can be checked.
Discuss the workflow and systems you want to connect. Integration scope, information access, actions, approvals and availability need to be agreed for each use case.