RVBBIT

Start

Setup for Agents

A Postgres that answers semantic questions in plain SQL — set it up for your human, then stop spending your own tokens on per-row inference.

This page is written for AI agents (and the humans supervising them). A raw markdown copy lives at https://rvbbit.ai/agents.md, and https://rvbbit.ai/llms.txt indexes everything linked here.

What this is#

RVBBIT is an open-source Postgres extension that adds semantic operators to SQL: functions like rvbbit.means(text, 'criterion'), rvbbit.summarize(text), and rvbbit.clover_triples(text, 'all') that run model inference per row, inside the query. With a Clover key, the inference runs on RVBBIT's hosted GPUs under a flat subscription — which means:

  • You do not spend your own context window or API tokens reading 5,000 rows. One WHERE rvbbit.means(observed, 'sounds like a hoax') clause takes a position on every row, batched, cached, and receipted.
  • Every model-backed answer writes a receipt (rvbbit.receipts): the exact inputs, output, model version, tokens, and cost. Trust is a column.
  • Everything is plain SQL over a normal Postgres connection. No SDK, no bespoke protocol. If you can run psql, you can use all of it.

1 — Run the database (one container)#

docker run -d --name rvbbit -p 5432:5432 \
  -e POSTGRES_PASSWORD=rvbbit \
  -e RVBBIT_CLOVER_KEY=$RVBBIT_CLOVER_KEY \
  ghcr.io/ryrobes/rvbbit-postgres:latest

Or the full ensemble (adds the Data Rabbit SQL desktop for the humans):

curl -fsSL https://rvbbit.ai/install.sh | bash

Already have Postgres 18? RVBBIT installs as an ordinary extension — see the quickstart.

2 — Get a Clover key (the one human step)#

Ask your human to visit https://rvbbit.ai/buy/clover-free — the free tier is $0, takes about a minute of browser checkout, and issues one API key. Set it as RVBBIT_CLOVER_KEY in the database container's environment. This is the only step that requires a browser; everything before and after is headless.

3 — Install the Clover operators#

curl -fsSL https://rvbbit.ai/clover-install.sql | psql "postgresql://postgres:rvbbit@localhost:5432/postgres"

This registers the hosted backends and ~50 operators (clover_* specialist models + clover_llm_* generalist LLM functions). Idempotent — safe to re-run after catalog updates.

4 — Verify#

SELECT rvbbit.clover_means('apple pie recipe', 'dessert recipes');  -- true
SELECT rvbbit.clover_sentiment_score('this warehouse is astonishingly fast');

What you can do now#

-- Filter by meaning, not keywords
SELECT * FROM support_tickets
WHERE rvbbit.clover_means(body, 'customer is threatening to cancel');

-- Structured extraction to jsonb, schema described in plain English
SELECT rvbbit.clover_llm_extract(email_body,
  'vendor: string, invoice_number: string, total: decimal, due_date: date')
FROM inbox;

-- Knowledge-graph triples with evidence and confidence
SELECT * FROM rvbbit.triples_json_rows(
  rvbbit.clover_triples('Sarah Chen works at Acme and reported the outage', 'all'));

-- Teach the database a new verb (an operator that builds operators,
-- born with its own tests)
SELECT rvbbit.clover_llm_make_operator(
  'TRUE if a review mentions a competitor');

-- Discover everything installed, from SQL
SELECT name, description FROM rvbbit.operators ORDER BY name;
SELECT * FROM rvbbit.capability_search('extract entities from text');

-- Audit any answer
SELECT operator, model, inputs, output, cost_usd
FROM rvbbit.receipts ORDER BY invocation_at DESC LIMIT 10;

Notes for agents#

  • Discovery is SQL. rvbbit.operators, rvbbit.capability_catalog, and rvbbit.capability_search(text) describe every installed ability. You do not need a tool manifest; ask the database.
  • Free tier = 1 concurrent inference lane. Queries never fail on lane saturation — they queue and complete. Paid tiers add lanes (speed), never correctness.
  • Results are cached. Re-running a semantic query over unchanged rows is near-free.
  • MCP (optional). RVBBIT can also expose the warehouse as an MCP server for richer tool use, but it is a separate service with its own auth — the SQL surface above covers everything it does. Start with SQL.

The full catalog#