Guides¶
Task-focused walkthroughs for operating and extending a Latence stack. New to the framework? Start with Getting Started and Concepts & Architecture first.
Operate a stack¶
-
Run & validate a stack Assemble a full-stack Pipeline config and validate it end-to-end with
latence stack validate. -
Incremental ingest Document churn without a full rebuild:
latence delta/retract/purge, O(changed) extraction, exact corpus-level recompute — equivalence to a full rebuild proven on real data. -
Tune a stack per device Auto-tune batch sizes and device knobs per host with
latence tune. -
Run a bake-off Compare Providers on a Capability — quality, throughput, license, and cost in one table.
-
The e2e campaign wizard
latence-campaign: gated, resumable benchmark stages from corpus acquisition through the full pipeline to measured retrieval — byte-identical configs, license lanes enforced.
Deployment & serving¶
-
Serving LightOn-OCR with vLLM The production OCR path: serve
parser.lighton_vllmbehind an OpenAI-compatible endpoint. -
GPU testing guide Run the whole learned/GPU system yourself, validated on a CUDA pod.
Integrate¶
-
Retrieval-engine input specification The exact contract a downstream retrieval engine consumes from a Latence run — the shape of the RAG-ready corpus and knowledge graph.
-
Graph-rescue retrieval The production graph-at-query-time architecture: base dense+BM25, a policy-gated graph lane that appends (never fuses), and a token-budget selector — measured defaults, zero LLM calls online.
-
Serving with Qdrant The verified Export → collection → served-queries path: pinned compose,
distance: dot,hnsw_ef: 512— recall parity with brute force measured on the full query sets. -
Embedders The opt-in RAG-corpus vector augmentation and how it plugs into Export.
Enterprise readiness¶
"Enterprise-ready" is not a marketing claim in Latence — it is the Provider Conformance
gate (Authoring a Provider §5) plus the declared, license-verified
ProviderProfile on every Provider
(ADR-0036). Together they mean: every
Provider fails gracefully with typed errors, leaks no PII, honors its device declaration,
and carries a verified weights-and-code license
(ADR-0012). Each run's
Quality Report folds the
conformance result in, so the "AI-ready" claim is substantiated per run, not asserted.