RVBBIT

Start

Overview

What RVBBIT is, how the pieces fit, and how to read the docs.

RVBBIT is a Postgres extension for SQL-native semantic work. It lets you keep ordinary tables, queries, and operational habits while adding model-backed operators, embeddings, knowledge graph workflows, and optional storage acceleration and routing.

The extension is not only a columnar acceleration experiment. The storage layer is useful, but the center of gravity is SQL as the control plane for AI and workflow execution.

Mental Model#

Think about RVBBIT in four layers:

Layer Job
Postgres heap The durable source of truth and universal fallback.
Semantic SQL Operators, embeddings, KG helpers, MCP calls, model costs, and receipts.
Acceleration registry Optional: registered heap tables get rebuildable columnar files beside them — Parquet, Vortex, Lance, hot memory, layout variants, and time travel — plus multi-engine routing (DataFusion, Duck, GPU GQE).
Warren Capability/runtime nodes for managed sidecars and workflow execution.

You can use semantic SQL without registering a single table for acceleration. You can accelerate tables without using a single semantic operator. The pieces are designed to compose, not to force one deployment shape.

Two more pieces round out the product:

  • The Docker ensemble. For v1, RVBBIT is distributed as a Docker Compose stack — Postgres 18 with the extension preinstalled, the Duck/Vortex worker, Data Rabbit, and a Warren agent. Treat it like a database, not like software you install into Postgres by hand. See Quickstart.
  • Data Rabbit, the desktop UI. A fast SQL desktop for any Postgres that lights up with cockpits for the whole rvbbit surface — operators, receipts, routing, the knowledge graph, cubes, metrics, and more.

What Is Novel#

The center of the system is not "an LLM function in SQL." The more interesting piece is that model calls, tools, workflow steps, graph memory, retrieval, and cost accounting can share one SQL-visible control plane:

  • Cascades keep multi-step model workflows behind typed SQL functions.
  • MCP tools can be called directly, joined as row sources, or used as Cascade steps.
  • Semantic operators cover retrieval, clustering, classification, deduplication, novelty detection, extraction, and evidence snippets. The core extension seeds LLM-backed means(), about(), classify(), and extract(), while capability packs add local-specialist versions of those names plus pack-only operators like extract_pii() (for example, reranker-backed means() / about() ship from the BGE reranker packs).
  • Receipts and cost ledgers make semantic calls inspectable after the fact.
  • The acceleration registry can accelerate analytical tables without changing the heap fallback contract.

What The Docs Optimize For#

Each topic starts with the path you should use first, usually a short SQL example. The same page then goes deeper into catalog tables, knobs, observability, and edge cases.

This is deliberate. RVBBIT has a lot of sharp, interesting machinery, but most users should only need the first screen of a page to get moving.

Good First Reads#

  1. Positioning for the product shape and mental model.
  2. Quickstart for the Docker ensemble, semantic SQL, and your first accelerated table.
  3. Examples for compact SQL snippets across the system.
  4. Semantic SQL for user-defined model-backed operators.
  5. Semantic Functions for built-in retrieval, classification, clustering, extraction, and evidence primitives.
  6. Cascades for multi-step operator workflows with gates, takes, validation, retries, and receipts.
  7. MCP Servers for bringing external tool ecosystems into SQL.
  8. Receipts And Costs for audit and cost accounting.
  9. Storage Acceleration for the acceleration registry and the heap fallback contract.
  10. Routing And Training for how RVBBIT chooses an execution path (rules and shadow observation today) and trains models.
  11. Data Rabbit for the desktop UI over all of it.
  12. Operations for test, benchmark, and observability habits.