Skip to content

System overview

Partial

The diagram reflects current processes; specialist legal features and the full agent platform remain incomplete.

Code and tests

Kesita separates private product work from the shared legal corpus. The diagram shows the implemented process and database boundaries; it does not claim that every component is deployed in Azure.

flowchart TB
  Browser["Browser"] --> App["Next app: sessions, routes and repositories"]
  App --> API["Product API"]
  App --> Product[("Product PostgreSQL")]
  API --> Product
  API --> Commands["Committed workflow commands"]
  Commands --> Worker["Product worker with DBOS"]
  Worker --> Product
  Worker --> Engine["Document jobs and documents-go inside worker"]
  Engine --> Parser["Document parser process"]
  Engine --> Blobs[("Private blobs")]
  Worker --> Gateway["Model gateway inside worker"]
  Worker --> Core["Legal Core API"]
  Core --> Legal[("Legal PostgreSQL and originals")]
  Corpus["Corpus worker"] --> Legal

The Next app still queries product PostgreSQL through server-side repositories, including onboarding, invitations and administration. Product API owns the Go workspace endpoints; it is not yet the exclusive product-data boundary. A dedicated document service is also not implemented: documentjobs and documents-go execute inside the product worker, while parsing is a separate process.

This matters when tracing authorization: inspect the actual app repository or Go handler and its database role rather than assuming every browser request passes through Product API. Services & ownership maps those paths. Identity & access explains the enforcement points.

MCP, specialist legal operations and Latera remain target capabilities. Their absence must not be inferred away from an architecture arrow. See the status notes on external integrations and Latera.

Legal Core now uses Qdrant for semantic/hybrid candidates, while PostgreSQL retains canonical content and relationships. The simplified diagram above shows canonical ownership; the additional legal-qdrant and legal-search-projector processes are wired in Compose. Storage & indexes owns the minimal payload, rebuild publication and remaining synchronization limits. No direct product-to-Qdrant access is introduced.