OpenLi OpenSOP turns your policy library into an agentic copilot — operators ask in plain English, the agent grounds itself in your SOPs, pulls live data, and returns one structured action with inline citations and hard guardrails on PII, contradictions and low-confidence retrieval.
Built on Claude-Code-style 3rd-generation autonomous agents — paired with a deterministic RAG agent for the boundary cases where structure beats autonomy.
Variance is HIGH (qty_var=20.0%, eta_var=10d). HIGH never auto-actions; matrix routes Womenswear / 200k+ to the Director of Merch role.
Audit-defensible recommendations. Hard PII guardrails. Run on any LLM provider.
Planners ask in natural English; the agent grounds itself in your SOPs, pulls live PO + forecast data, and returns one structured action with citations.
pgvector HNSW cosine over every SOP doc. Every recommendation cites the source document and section — audit-defensible by construction.
PII never surfaces (role-only escalation). Contradictions force escalate with low confidence. Low retrieval confidence refuses to invent policy.
Toggle between OpenLi Codex agent, Claude Agent SDK runner, or our dedicated RAG agent — per call. Same recommendation contract from all three.
amend · split_child_po · firm_planned_order · raise_backorder · escalate. Closed schema. Out-of-set values coerced. Safe to wire into automation.
Each business unit gets its own SOP corpus inside an isolated workspace. Upload, scan, or git-import. Re-index on demand.
Plus a deterministic RAG fallback for the safety boundary cases.
Chunk + embed + retrieve. Useful for Q&A; can't take structured actions.
Hard-coded DAGs. Brittle. Hard to evolve when SOPs change.
Claude-Code-style — read files, plan steps, call tools. We layer hard schema + safety guardrails on top.
Drag policy PDFs/MDs into the tenant workspace. Click Reindex. pgvector embeddings + HNSW index in seconds.
Expose your PO / forecast / ticket lookup as simple Python or HTTP tools. The agent decides when to call them.
OpenLi Codex (OpenAI Codex SDK), Claude Agent SDK, or our dedicated RAG agent. Toggle per call — same output contract.
PII scrub, contradiction detection, low-confidence-refuse — layered on every path. Eval harness ships in the repo.
ASOS Merch is the reference. Procurement, Ops, Compliance, Clinical — same RAG + tools + guardrails pattern, different corpus.
PO variance, child-PO split, retail-vs-wholesale allocation, backorder thresholds, escalation matrix.
Three-way match exceptions, supplier risk SOPs, payment-term overrides — same agent pattern, different corpus.
Incident triage, SOX control reviews, internal-audit SOPs. Auto-classify, cite policy, escalate to role.
Patient-pathway SOPs, drug-interaction overrides, adverse-event triage — same RAG+tools+guardrails shape.
AGPL-3.0 community license. Commercial license for organisations with revenue ≥ £250k or air-gapped deployments.
Open the live demo — ask a question about a problem PO, watch the agent ground itself in the SOPs, pull live data, and emit a structured recommendation with citations.