Clover · operator reference · new to the term?
clover_* function below is plain SQL after curl -fsSL https://rvbbit.ai/clover-install.sql | psql. The diagrams are the real execution DAGs from the installer — arguments flow into hosted model steps (and deterministic code gates), and out as a typed SQL value. Receipts record the exact model version for every call.Before the list — the one weird argument
model_blob_b64 — you hold the model, not us.clover_fit and clover_anomaly_fit train a model on your rows and return it to you as a base64 blob inside the JSON result (->>'blob_b64'). Nothing is stored server-side — the model is just data in your database, like any other value. Operators that take model_blob_b64 are asking for that string back. The usual shape is one statement, fit feeding predict through a CTE:WITH model AS (
SELECT rvbbit.clover_fit(
'classifier',
'[[1,1],[2,1],[1,2],[8,9],[9,8],[9,9]]', -- features
'["small","small","small",
"big","big","big"]' -- labels
) AS m
)
SELECT rvbbit.clover_predict(
m->>'blob_b64', -- the model, straight back in
'[[9,9]]' -- rows to classify
)->'predictions'->>0
FROM model; -- → big-- Train once, store the blob like any value:
CREATE TABLE churn_model AS
SELECT now() AS trained_at,
rvbbit.clover_fit('classifier',
f.features, f.labels) AS m
FROM my_training_set f;
-- Reuse it anywhere, forever, no refitting:
SELECT rvbbit.clover_predict(
(SELECT m->>'blob_b64' FROM churn_model
ORDER BY trained_at DESC LIMIT 1),
c.features)
FROM new_customers c;Same pattern for clover_anomaly_fit → clover_anomaly_score, and clover_explain takes the same blob (plus its sha256, also in the fit result) to produce SHAP attributions.
Encoder specialists
Fit an anomaly detector on numeric rows (client-held model)
Returns your model. The fitted model comes back IN the result — result->>'blob_b64' — nothing is stored server-side. Hand that string to clover_anomaly_score / clover_explain, usually via a CTE (worked example ↑), or save the row in a table to reuse it forever.
Score rows for anomalousness against a fitted detector
model_blob_b64 = a model you fitted earlier. It's the blob_b64 field returned by clover_anomaly_fit — chain them in one statement with a CTE (worked example ↑) or read it back from wherever you stored it.
Zero-shot classification with the winning label and every candidate score
Cluster a JSON array of texts into assignments and representative groups
Parse a document into Markdown, typed blocks, tables, cells, hierarchy and source geometry
Reusable Arctic text embedding vector for KNN, similarity and clustering
SHAP feature attributions for a fitted TabPFN or anomaly model
model_blob_b64 = a model you fitted earlier. It's the blob_b64 field returned by clover_fit / clover_anomaly_fit — chain them in one statement with a CTE (worked example ↑) or read it back from wherever you stored it.
Entity extraction (GLiNER) over your types
Fit a TabPFN model on tabular rows — the fitted model is returned to you
Returns your model. The fitted model comes back IN the result — result->>'blob_b64' — nothing is stored server-side. Hand that string to clover_predict / clover_explain, usually via a CTE (worked example ↑), or save the row in a table to reuse it forever.
Forecast a numeric series N steps ahead — median + quantile bands
Forecast named related series with historical and known-future business drivers
Reusable SigLIP2 embedding for an image or text description
Language code plus confidence for thresholded routing
Full 3-way entailment, neutral, and contradiction scores
PII detection preset (person, email, phone, ssn, ...)
Predict with a clover_fit model — classifications or regressions
model_blob_b64 = a model you fitted earlier. It's the blob_b64 field returned by clover_fit — chain them in one statement with a CTE (worked example ↑) or read it back from wherever you stored it.
Extract (subject, predicate, object) relation triples from text
Relevance score 0..1 of text to criterion
Sentiment label + score for a text
Score a long time series for anomalies with CPU-side TSPulse
Create a semantic TSPulse embedding for the most recent 512 points
Fill missing points in a time series with TSPulse reconstruction
Research the public web with cited claims, typed knowledge, tenant-scoped snapshots, and exact evidence lineage
Research with one bounded follow-the-evidence pass to close material gaps
Fetch a public HTTP(S) page or document and return Markdown, metadata, provenance, and extraction diagnostics
Refresh stable public URLs, reuse unchanged evidence without another synthesis, and report source and claim changes
Composite cascades
Zero-shot classification over your labels
TRUE if two statements contradict (NLI 3-class)
TRUE if premise entails hypothesis (NLI)
Similarity of two images, or an image and a text description
ISO language code of text
TRUE if text semantically matches a criterion (cross-encoder)
Full moderation category scores
OCR a document to plain text through the structured Granite-Docling parser
Continuous sentiment in [-1, 1]
Compare two time-series shapes with TSPulse semantic similarity
Embedding cosine similarity of two texts
TRUE if text is toxic
Transcribe audio to text (Whisper large-v3-turbo)
Fetch a public HTTP(S) page or document and return its cleaned Markdown
LLM operators
Rewrite text with all PII redacted ([NAME], [EMAIL], ...)
Apply an instruction/question to a text — answers only from the text
One-shot ask against the hosted generalist
Canonical form of a value (NYC -> New York City)
Synthesize the shared consensus across a JSON array of findings
TRUE if two texts oppose each other on the same subject
Messy text to ISO 8601 date, or NULL
Freeform text to a JSON object per a schema you describe
Detect logical fallacies — JSON array of {fallacy, explanation}
Repair malformed values given a type hint (gmial.com -> gmail.com)
TRUE if premise implies conclusion — incl. the quantifier logic fast NLI misses
TRUE if text semantically matches a criterion — LLM-judged for inputs that need thought
Merge duplicate JSON records into one canonical golden record
TRUE if two values identify the same real-world entity
Judge text against any English-phrased criterion, 0-1
Rewrite an argument in its strongest, most defensible form
How strongly evidence supports a claim, 0-1
Extract a normalized chronological event timeline from text
Translate to any language — returns only the translation
TRUE if a value satisfies an English-phrased rule
Extract knowledge-graph triples from text as strict JSON — same contract as the built-in rvbbit.triples, so every KG surface (data_crawl, Document Brain, Scry) can run on it