What I am building
I separated SovereignFlow from LocalAI-RAG after seeing how deeply code-analysis assumptions shape a system. Removing words like “C#” or “repository” does not make an engine general-purpose, so I am building a neutral DocumentChunk model, ingestion and retrieval contracts, and pipeline execution that does not depend on document subject matter. The repository contains infrastructure code, tests, migrations, and documentation.
Neutral core and domain applications
The engine owns workflow execution, retrieval, audit, policies, and APIs. Domain applications own their own importers, business rules, schemas, and prompts. A tariff-classification application should not force the core to understand tariffs, just as code analysis should not define every RAG concept.
Ingestion and data consistency
An import stores a source version, chunks, idempotency key, and indexing job in PostgreSQL. Vectors go to Weaviate, and the current-version pointer changes only after successful indexing. The documentation also describes a separate large-JSONL importer with resumption and consistency checks.
Audit and security contracts
Pipeline executions and steps are recorded in PostgreSQL audit data. Authorization resolves verified permissions into capabilities and pipelines rather than accepting arbitrary pipeline names from clients. Model endpoint selection is explicit; errors do not silently switch providers. These are architectural decisions I validate through contracts and tests.
Project scope
- DocumentChunk and domain/tenant-neutral metadata
- Versioned YAML action behavior and step-graph validation
- PostgreSQL execution audit, migrations, and durable ingestion jobs
- Weaviate retrieval with ACL/classification checks before context assembly
- Explicit authorization policies and local/external model selection
- Document chunk relationships and bounded graph expansion