Getting Started¶
Latence produces two AI-ready deliverables from a folder of messy documents — a RAG-ready corpus and a knowledge graph, both as portable files. This section gets you from an empty environment to those outputs.
The one-command path¶
Everything ships CPU-first and offline by default (ADR-0007), so you can run a full pipeline on a laptop with no GPU and no network:
# Install from a clone — nothing is published to PyPI (see the note below).
git clone https://github.com/ddickmann/latence
cd latence
uv sync
# 1. A guided wizard emits ONE opinionated, quality-baked stack config (ADR-0041).
uv run latence setup
# 2. Run that config end-to-end over your documents.
uv run latence process --config stack.yaml --input ./my-documents --out ./corpus
Why a clone, and not pip install
Nothing is published to PyPI: the distribution channel is the signed GitHub Release
(ADR-0062,
whose rationale is superseded and whose decision is reopened — not reversed — by
ADR-0063). A pip install latence-core today
fails with ERROR: No matching distribution found. uv sync installs the whole
CPU-first workspace from the clone, which is what every command on this page assumes;
the quickstart walks the same install step by step.
latence process runs the wizard-produced stack and writes the RAG corpus + knowledge
graph, plus a first-class Quality Report
substantiating the "AI-ready" claim.
# 3. Read the results — a read-only, offline, localhost console over what the run wrote.
# `latence-console` is already in the environment `uv sync` built.
uv run latence-console ./corpus
latence-console serves the run list and each run's Quality Report as headline quality
verdicts a non-engineer can read, not raw JSON. It writes nothing, holds no state, and makes
no network requests — see Read the results.
Run it stage by stage¶
To understand what each Stage does — and to swap in a learned or GPU Provider and see the quality lift — follow the walkthrough:
-
Fresh-pod walkthrough From a bare GPU pod to AI-ready data over a mixed-file corpus with the enterprise-SOTA pipeline (served-vLLM OCR, GLiNER2 extraction, the GLinker entity-linker, Granite r2 embeddings) — every Stage in isolation, then the full end-to-end run.
-
GPU testing guide Run the whole learned system yourself: the GPU Providers, validated on a CUDA pod.
The core CLI¶
The latence CLI (installed with latence-core) is the entry point for every workflow:
| Command | What it does |
|---|---|
latence setup |
Guided wizard → one opinionated, quality-baked stack config (ADR-0041). |
latence process |
Run a full stack config end-to-end over a document folder. |
latence run |
Run an explicit Pipeline file. |
latence stack validate |
Validate a stack config end-to-end (see Run & validate a stack). |
latence bake-off |
Compare Providers on a Capability (see Run a bake-off). |
latence tune |
Auto-tune a stack per device (see Tune a stack per device). |
latence delta |
Run an incremental corpus delta (ADR-0018). |
latence retract / latence purge |
Soft-tombstone or hard-erase documents (GDPR). |
The read-only run console ships as its own package (pip install latence-console) so the core
CLI stays dependency-thin:
| Command | What it does |
|---|---|
latence-console <location> |
Serve the read-only run console over a Storage location (Read the results). |
Where to go next¶
- Want to show a run to a non-engineer? Read the results — the run console.
- New to the model? Read Concepts & Architecture — the Capability / Provider / Stage / Runner / Storage vocabulary the rest of the docs assume.
- Want to see what you can plug in? Browse the Provider catalog.
- Building your own Provider? Start with Authoring a Provider.