HarborRAG

What is HarborRAG?

HarborRAG is a provider-agnostic Retrieval-Augmented Generation framework for engineering knowledge. It uses ports and adapters so source systems, document formats, model providers, persistence backends, and operator surfaces can change independently.

The first public alpha release is HarborRAG 2.0.0 Alpha 1 (Python package version 2.0.0a1), dated August 27, 2026. It continues the project lineage of Qdrant Loader as a breaking rename and a provider-neutral architectural expansion; legacy import paths and package names do not carry over.

Capability status

HarborRAG is under active alpha development. The distinction between implemented adapters and the default application path matters:

Layer Status
Core domain and schemas Implemented provider-neutral document, retrieval, model, storage, tenancy, security, and telemetry contracts
Source connectors Implemented for local files, GitHub, Confluence, Jira, and SharePoint
Document parsers Implemented for common text, structured, Office, image, ebook, and PDF formats
Model adapters Implemented chat, embedding, and reranking clients with validation, routing, retries, caching, budgets, and telemetry boundaries
Repository adapters Implemented backends for Qdrant, FalkorDB, Redis, PostgreSQL, SQLite, S3, filesystem, and memory
Engine and runtime Postgres-authoritative ingestion, durable Temporal workflows, immutable artifacts, hybrid vector projection, graph projection, retrieval, chat orchestration, stored prompts, and production dependency composition are implemented
CLI doctor, one-shot chat, Temporal-backed ingestion control, and authoritative retrieval commands
HTTP API FastAPI factory serving a public /v1 contract (ingestion, retrieval, chat, agent, admin) plus an operational /api/v1 surface (liveness, readiness, Prometheus metrics, console routes)
MCP Audited, policy-bounded FastMCP transport exposing four tenant-scoped retrieval tools (vector_search, graph_triplet_search, graph_path_search, graph_subgraph_search) over stdio or bearer-authenticated loopback HTTP with a local status/configuration UI
Temporal deployment PostgreSQL-backed local development stack; production deployment remains application/operator work

Why the package boundaries exist

Engineering RAG combines several failure domains: source authentication and pagination, hostile or malformed documents, model routing, tenant-aware persistence, retries, orchestration, and public interfaces. Keeping them separate provides three practical benefits:

The package dependency direction is checked in CI. See Architecture for the exact graph.

Intended use today

Use the repository to develop and test adapters, run the PostgreSQL-backed Temporal ingestion stack, and execute opt-in real-system acceptance checks in a controlled environment. Internet-facing production deployments still require operator-owned identity policy, TLS/network controls, secret delivery, backup/restore, resource limits, and alerting.

Next steps