locus-etl 0.0.2

Locus — turn any unstructured corpus into validated, source-grounded tabular data. CLI: `locus`.

Locus

Turn any unstructured corpus into validated, source-grounded tabular data — ready to feed an LLM.

Locus packages data operations as reusable, versioned images. You pull an image, point it at your own data, run it locally, and get a clean table where every cell carries its source location and a faithfulness score. Images compose into pipelines, and you can publish your own to Locus Hub (public or private).

Install

pip install locus-etl          # the CLI command is `locus`

Optional extras: pip install "locus-etl[pdf,serve,llm,oci]" (PDF parsing, Hub/result UI, LLM engine, OCI registry).

Quickstart

locus catalog list                  # see the official image catalog
printf 'name,amount\nAcme,100\nGlobex,200\n' > data.csv
cat > locusfile.yaml <<EOF
image: doc-to-tables
source: { type: files, path: ./data.csv }
EOF
locus run locusfile.yaml --export out.csv   # grounded table + _lineage column
locus run locusfile.yaml --serve            # preview UI with per-cell provenance
locus hub                                    # browse the catalog in a local web UI

Architecture (layered)

The diagram is generated from docs/generate_architecture_diagram.py (PNG + SVG in docs/).

Layer summary

Layer	Spec	Responsibility
Layer 1 — Engine	unstructured-to-tabular-etl	Raw corpus -> validated, source-grounded table. Connectors, parsing, extraction, cleaning, the cell-level grounding/faithfulness contract, review. Embedded inside every image.
Layer 2 — Runtime	locus-image-runtime	Packaging, CLI, Locusfile, image pull, multi-image composition (DAG), typed stage interchange, cross-stage provenance, serve/export, and publishing to Locus Hub.

Key properties

• Local-first. Default runtime is a plain Python process — no daemon, no Linux VM. Docker is an optional backend.
• Privacy is explicit. Deterministic engine keeps data local; the LLM engine activates only when you add a key, with a consent notice before any data leaves.
• Trust travels with the data. Provenance and faithfulness survive every pipeline stage, from extraction through merge and redaction.

Documentation

Detailed design lives in the spec documents:

• Layer 1 engine — .kiro/specs/unstructured-to-tabular-etl/requirements.md
• Layer 2 runtime — .kiro/specs/locus-image-runtime/requirements.md
• Image catalog (planned images + build order) — .kiro/specs/locus-image-runtime/image-catalog.md
• Architecture & decision log — .kiro/specs/unstructured-to-tabular-etl/architecture-notes.md

Status

Layer 1 engine: feature-complete. Layer 2 runtime: feature-complete (all 12 build stages done). Raw corpus → validated, source-grounded table with cell-level provenance; a deterministic default engine and opt-in guardrailed LLM engine; cleaning/dedup; human-in-the-loop review; file/HTTP/REST/SQL connectors with CSV/PDF/HTML/records parsers and DataFrame/Parquet/SQL emitters. The locus CLI runs single images and multi-stage pipelines (typed DAG with static type-check + cross-stage provenance), builds/publishes/pulls images via a local registry, and serves a local result UI with the provenance viewer. 208 tests, CI on Python 3.11/3.12 (ruff + mypy strict + pytest).

pip install locus-etl        # CLI command is `locus`; extras: [pdf] [llm] [serve] [oci] ...
locus init                   # gitignore .env
locus catalog list           # see the official image catalog
locus run locusfile.yaml     # run a pipeline, get a grounded table
locus run locusfile.yaml --serve --port 8080   # preview UI with provenance
locus hub                    # browse the image catalog in a local web UI
locus build / push / pull / search / inspect   # image lifecycle

Remaining work is the official image catalog and a hub-side discovery index. The OCI/Harbor registry backend is implemented (OrasImageStore): set LOCUS_REGISTRY (and optionally LOCUS_NAMESPACE) to push/pull/inspect against Harbor, GHCR, ECR, or any OCI registry; otherwise a local filesystem registry is the zero-config default.

from locus_engine import (
    Pipeline, PipelineConfig, PluginRegistry,
    FileConnector, CsvParser, Connector, Parser, SourceRef,
)

registry = PluginRegistry()
registry.register(FileConnector(), Connector)
registry.register(CsvParser(), Parser)

config = PipelineConfig.load({"source": {"type": "files", "path": "./data"}})
pipeline = Pipeline(config, registry)

out = pipeline.run([SourceRef(uri="./data/invoices.csv", kind="file")])
frame = pipeline.emit(out)          # pandas DataFrame with a _lineage column
print(frame)

Contributing

See CONTRIBUTING.md.

License

MIT © 2026 Dibae101