Let’s talk
Internal R&D

DecisionTrace

Auditable Decisions Across a Graph, a Relational Store and a Vector Index

Internal prototype — no customer

A prototype for auditable decision intelligence that answers from three stores at once: a graph for relationships, PostgreSQL for metadata and logs, and a vector index for semantic similarity.

Industry
Internal / BrainGenz R&D
Location
Internal
The problem

Decision auditability requires answering three different kinds of question about the same event: what it was connected to, what it resembled, and who recorded it. Each maps to a different storage model.

  • Relationship questions need graph traversal
  • Similarity questions need vector search
  • Provenance and audit questions need relational integrity
  • One store handles at most one of these well
  • Running three stores adds real operational cost
What we built

A containerised stack running a graph store, a relational store and a vector index together, with an API layer that routes each question to the appropriate path or combines them.

  • Use a graph database for relationship and dependency context
  • Use PostgreSQL for metadata, provenance and decision logs
  • Use a vector index for semantic similarity across past decisions
  • Orchestrate all three with health checks and persistent volumes
  • Route each query to the appropriate store, or combine results

How it was built

Structural and semantic questions need different stores

Asking which decisions depended on a given input is a graph traversal. Asking which past decisions resemble this one is a vector search. Asking who approved it and when is a relational lookup. Forcing all three into one store means two of them run badly.

Three stores, one query path

The prototype runs a graph database, PostgreSQL and a vector index together under container orchestration with health checks and persistent volumes, and routes a question down the appropriate path or combines them. The value being tested is whether hybrid retrieval in one system produces better reasoning traces than any single store.

Where it stands

A prototype built to explore the architecture. There is no customer and no production deployment.

Before
    After
    • Hybrid structural and semantic retrieval from one interface
    • Prototype only, with no production deployment
    Stack
    Stores
    Neo4jPostgreSQLWeaviate
    Orchestration
    Docker ComposeHealth checksPersistent volumes
    Backend
    API layer over three query paths
    Talk about work like thisBack to the work