Reference · Enterprise CLI

Enterprise CLI

Enterprise adds source inspection, record preparation, A–E experiment control, final approval, and agent query routing to the Rust corec command.

Packaging and invocation

Customers invoke corec. The licensed package includes a private Rust executable that corec discovers as its Enterprise implementation. COREC_EE_BIN can select a non-default extension location for development or controlled packaging.

shellCoreCortex Enterprise

corec strategy list
corec corpus classify --help
corec schema advise --help
corec dataset prepare --help
corec design --help
corec manifest --help
corec query --help

Inspect the customer strategy inventory

The base Enterprise build reports the four Core chunkers and an empty optionalPacks list. An enablement adds the packs selected for the customer's datasets. Dataset materialization checks this inventory, and its receipt binds each Stage B configuration to an available processor.

Command	Purpose	Primary output
corec strategy list	Show Core strategies and optional packs enabled in this build	StrategyInventory JSON
corec strategy describe <pack-id>	Inspect compatibility, licensing, record types, processors, versions, and configuration schema	ChunkingStrategyPackDescriptor JSON

Understand the source corpus

All three commands treat the source as immutable. Exploration hashes distinct values and omits examples by default. Enrichment adds explicit derived fields and keeps their origin visible.

Command	Required inputs	Primary output
corec corpus classify	--input JSONL\|Parquet, --output directory	classification.json and candidate dataset partitions
corec corpus explore	--input JSONL\|Parquet, --output file	FieldProfile JSON; --include-examples is optional
corec corpus enrich	--input structured JSONL\|Parquet, --output JSONL	Deterministically enriched rows with provenance

JSONL and Parquet support

Enterprise reads source exports in JSONL or Apache Parquet. Classification, exploration, deterministic enrichment, schema advice, and canonical dataset preparation use the same Rust source reader for both formats. Parquet values and explicit nulls are preserved through Arrow record batches.

Reviewable and portable outputs remain JSON or JSONL: canonical records, schema registries, golden judgments, ranked results, misses, and evaluation reports. corec dataset prepare-batch consumes canonical records.jsonl and emits bounded API request parts. This gives analysts inspectable evidence while allowing source teams to deliver efficient columnar exports.

shellCoreCortex Enterprise

corec corpus classify --input exports/documents.parquet --output engagement/00-classification
corec corpus explore --input exports/documents.parquet --output engagement/02-profile.json
corec schema advise --input exports/documents.parquet --name reference-documents --output engagement/03-advice.json
corec dataset prepare --input exports/documents.parquet --schema engagement/04-registered-schema.json --name reference-documents --provider customer-export --observed-at 2026-08-26T00:00:00Z --allow group:reader --output engagement/05-dataset

Stage	Accepted input	Produced artifact
Classify / explore	source.jsonl or source.parquet	Classification JSON, partitions, and field profile JSON
Enrich / advise	structured JSONL or Parquet	Enriched JSONL and pipeline advice JSON
Dataset prepare	source JSONL or Parquet	Canonical records.jsonl plus dataset.json
Prepare batch	Canonical records.jsonl	request.json parts and integrity receipts
Golden and eval	Reviewable JSONL	Fragments, judgments, results, misses, and reports

Advise and approve schema

Command	Required inputs	Primary output
corec schema advise	--input, --name, --output	PipelineAdvice with proposed identity, text, types, attributes, and transforms
corec schema register	--suggestion, --registry, --approved-by	Append-only registry revision; optional --output registered schema

Prepare and ingest datasets

Preparation requires an explicit access decision for every record set. It never infers reader permissions from customer fields. The resulting requests remain portable across adapter implementations.

Command	Key options	Primary output
corec dataset prepare	--input, --schema, --name, --provider, --observed-at, plus --allow repeated or --public	Canonical records.jsonl and dataset manifest
corec dataset chunk	--input, --output, --strategy, optional --configuration	Reviewable fragment JSONL and a hashed ChunkingRun receipt
corec dataset prepare-batch	--input, --name, --configuration-revision, --mapping-fingerprint, --output	Backend-neutral request parts, hashes, and prepared manifest
corec dataset ingest	Prepared batch plus canonical database and adapter settings	Durable application and projection receipt
corec dataset ingest-demo	--database-url, --config, --batch, --design-configuration, --output	Starter-only PostgreSQL and memory-projection receipt
corec validate-batch	Prepared request JSON	Offline public-contract validation

Create review candidates and run evaluations

The evaluators are public Core commands because their evidence contracts should remain independently inspectable. Enterprise prepares and coordinates the experiment workflow around them.

Command	Purpose	Gate evidence
corec golden generate	Build candidate judgments from exact materialized fragments	Stable fragment IDs; candidates remain unreviewed
corec eval retrieval	Score BM25 or supplied lexical, semantic, or hybrid rankings	Hit@K, misses, expected-empty behavior, configuration and projection-receipt binding
corec eval generation	Score answers, citations, facts, abstention, and supplied faithfulness	Quality, faithfulness, abstention, model and judge identities

Control stages A through E

record-result also accepts --converter-audit, --chunk-audit, --ingestion-evidence, --generation-report, and --latency-ms. Ingestion evidence is mandatory for A–C candidates and for D, where it must be the exact selected C receipt.

Command	Purpose	Evidence requirement
corec design init	Validate a declared search space and plan Stage A	--search-space, --output
corec design record-result	Attach one evaluator result to a planned experiment	--state, --experiment-id, --retrieval-report, --results; audits and receipts as applicable
corec design advance	Select the stage winner and plan the next experiments	Complete equal-sample results for the active stage
corec design audit-converters	Record concrete converter provenance and representative review	--fragments, --requested-converter, --reviewed-by, repeated --representative
corec design audit-chunks	Measure exact token and split distributions	--fragments, --output
corec design finalize	Bind the A–D optimum to full Stage E reports	Exact selected configuration, ready ingestion receipt, receipt-bound retrieval, and optional corpus report
corec design demo-stages	Exercise controller wiring with the shipped synthetic fixture	Artifacts are labeled synthetic_demo

Emit the certification manifest

Add --corpus-report for a multi-collection corpus. The result is certified only when every immutable gate passes; otherwise it is blocked and carries actionable reasons.

shellCoreCortex Enterprise

corec manifest emit \
  --classification <classification.json> \
  --sample-review <sample-review.json> \
  --profile <profile.json> \
  --advice <advice.json> \
  --schema-registry <schema-registry.jsonl> \
  --dataset <dataset.json> \
  --ingestion <applied-receipt.json> \
  --judgments <reviewed.jsonl> \
  --design-decision <design-decision.json> \
  --retrieval-report <report.json> \
  --generation-report <generation-report.json> \
  --certified-by <principal> \
  --retrieval-target 0.55 \
  --faithfulness-target 0.80 \
  --qualified-answer-target 0.80 \
  --output <manifest.json>

Route agent queries

Optional flags are --dataset for dataset scope, --filters for a JSON object, --global-search for an explicit global scope, and --output for durable JSON. The routing decision can select lexical, hybrid, or bounded graph followed by evidence retrieval. Relationship intent always carries requiresAgentReview: true.

shellCoreCortex Enterprise

corec query route \
  "Which authorities connect these two matters?" \
  --filters ./request-scope.json \
  --output ./routing-decision.json

Artifact and security rules

Input exports remain read-only and carry SHA-256 identity into derived artifacts.
Every command writes to an explicit file or directory and produces machine-readable JSON or JSONL.
Access uses explicit --allow principals or --public; customer fields never silently become permissions.
Secrets stay in ignored environment files or a secret manager.
Canonical records, reviewed judgments, and certification evidence remain independent of a vector or graph backend.

Continue readingAgent skills→