Capability Packs

Capability packs are how RVBBIT moves beyond a pile of SQL functions. A pack can install operators, backends, specialist runtimes, model metadata, smoke tests, or Warren deployment jobs.

Use packs when a feature needs assets and runtime shape, not only a single SQL function.

The friendliest install path is the Capabilities window in Data Rabbit — browse the catalog, click install, watch the deploy graph — which drives the same Warren SQL deployment documented below. The shell CLI is the development/repeatable-install path.

Local Pack Workflow#

capabilities/tools/rvbbit-capability list

capabilities/tools/rvbbit-capability scaffold \
  capabilities/packs/extract/gliner-medium-v2.1 \
  /tmp/rvbbit-gliner

Install a pack from the shell during development:

RVBBIT_DSN=postgresql://postgres:rvbbit@localhost:55433/bench \
capabilities/tools/rvbbit-capability install \
  capabilities/packs/extract/gliner-medium-v2.1 \
  --gpu

The command-line path is useful for development and repeatable installs.

SQL Deployment Through Warren#

Fresh extension installs seed rvbbit.capability_catalog, so a SQL client or UI can queue a catalog install without reading files from the server:

SELECT rvbbit.deploy_catalog_capability(
  catalog_id => 'extract/gliner-medium-v2.1',
  target_selector => '{}'::jsonb
);

For runtime capabilities, use a target selector that matches the Warren worker labels:

SELECT rvbbit.deploy_catalog_capability(
  catalog_id => 'runtimes/python-runtime',
  target_selector => '{"docker":true}'::jsonb
);

Warren workers claim jobs, materialize runtimes or model backends, and write status back to SQL. Queued jobs, target selection, logs, errors, smoke-test state, and generated SQL are all queryable.

Browse the static docs catalog at /capabilities, or query the live database:

SELECT id, title, operators
FROM rvbbit.capability_catalog
WHERE active
ORDER BY title;

What A Pack Can Include#

Asset	Example
Backend rows	A local embedding server or OpenAI-compatible local model.
Operators	Extraction, sentiment, classification, summarization.
Runtime manifests	Docker/Python/Rust sidecars needed by a specialist.
Cost policy	Free local GPU, fixed per-call, or model-rate estimate.
Smoke tests	SQL calls that prove the pack is usable.
UI metadata	Description, labels, requirements, and target compatibility.

Where The Familiar Operators Come From#

Some friendly names exist out of the box and some arrive with a pack — and a few exist in both forms. The core extension seeds LLM-backed means, about, classify, extract, summarize, sentiment, and triples as editable rows in rvbbit.operators, so they work before you install anything. Installing a capability pack adds local-specialist-backed operators that reuse some of those same names (a reranker model behind means/about, DeBERTa behind classify, GLiNER behind extract) and also adds names that are pack-only and never core-seeded (extract_pii, has_pii, semantic_score, semantic_matches, similar_to). When a name is exported by more than one competing pack, whichever you install (or install last) provides that backing.

Operator	Returns	Backing
means(text, criterion) (infix ~~? / MEANS)	bool	core seeds an LLM version; the BGE/MS-MARCO reranker packs (rerank/bge-reranker-base, rerank/bge-reranker-v2-m3, rerank/ms-marco-minilm-l6-v2) add a local reranker version
about(text, topic) (infix ~~% / ABOUT)	float8	core seeds an LLM version; same three rerank packs add a local reranker version
semantic_score, semantic_matches	float8 / bool	pack-only — the same three rerank packs (no infix form)
classify(text, categories), semantic_classify	text	core seeds an LLM classify; classify/deberta-v3-zero-shot and classify/deberta-v3-base-zero-shot add local zero-shot versions
extract(text, what)	text	core seeds an LLM version; extract/gliner-medium-v2.1 (GLiNER) adds a specialist version
extract_entities, extract_pii, has_pii	jsonb / text / bool	pack-only — extract/gliner-medium-v2.1 (GLiNER)
semantic_embed(text), similar_to(left, right)	jsonb / float8	pack-only — local embedding packs (embeddings/bge-small-en-v1.5, embeddings/bge-m3, embeddings/e5-small-v2)

Only about and means declare infix forms; semantic_score and semantic_matches are plain functions. The per-pack raw wrappers (for example rerank_bge_base_score, embed_bge_small, summarize_bart) keep their own prefixed names — the pack's friendly operators above are the multi-step Cascades built on top of them. For canonical retrieval, prefer rvbbit.embed / rvbbit.set_default_embedder(...) over the exploratory embed_* raw wrappers. See /docs/semantic-functions for the extension's own semantic primitives, /docs/predictive-models for the trained-model and tabular-foundation packs, and /docs/cascades for how multi-step operators are defined.

When To Use A Pack#

Good pack candidates:

• extractors such as GLiNER,
• local embedding models,
• domain-specific operator sets,
• MCP gateway bundles,
• specialist runtimes that need a worker process.

Avoid packs for one-off prompt experiments. Put those in normal operator catalog rows until the workflow is stable enough to distribute.

Observability#

SELECT *
FROM rvbbit.warren_jobs
ORDER BY created_at DESC
LIMIT 20;

SELECT name, effective_status, heartbeat_age, inventory
FROM rvbbit.warren_node_effective_status
ORDER BY last_heartbeat DESC NULLS LAST
LIMIT 20;

rvbbit.warren_node_effective_status is a view over rvbbit.warren_nodes that folds the reported status together with heartbeat freshness. Backend health for individual specialists comes from rvbbit.backends and rvbbit.backend_probe(backend_name); runtime registration lives in rvbbit.python_runtimes and rvbbit.mcp_gateways.

Warren is a runtime inventory: what can be installed, where it can run, what is currently deployed, and what failed.