OE knowledge corpus (advisory retrieval)
- Status
- contract + direction
- Conceptual layer
- ④ Decision
- Repo layer
- L4
knowledge-reasoning - Source
- architecture section 8, section 3.6.4
- Research
../../research/plant-efficiency-exploration-2026-09/20-oe-knowledge-retrieval.md- Normative
00-kernel.md— retrieved text never evaluates gates, invents ₹, or writes equipment.- Siblings
05-context-engineering.md·06-memory.md·11-models-and-seams.md·13-improvement.md
Purpose#
Plant memory answers what worked here. Detectors answer what is happening now. The OE knowledge corpusVersioned literature + sparse method ontology; retrieved as advisory rows only answers: for this class of situation, what methods does industrial practice use to drive efficiency?
Without it, dual-family models lean on weights and thin playbook bullets. That is not enough when Stamped wants excellence across energy, cost, time, continuity, and exceptions.
Decision#
| # | Topic | Decision |
|---|---|---|
| 1 | Gap | First-class OE knowledge corpus — separate from Hindsight, case library, and PSM |
| 2 | v1 retrieval | Sparse method ontology + vector retrieval (Vector-First or light Parallel Hybrid). Graph is methods↔scenarios↔signals, not plant topology |
| 3 | Partition | All hits → advisory ledger rows only |
| 4 | Authority | Advisory. Case library wins outcomes. Calculator owns ₹. Code owns constraints |
| 5 | Plant path | No Agentic GraphRAG; no unbounded research loops |
| 6 | Offline | Council / corpus curation may use heavier GraphRAG patterns to propose playbook bullets |
| 7 | Pins | Corpus version + embedding model in release lockfile |
Rejected: dumping books into the system prompt; using PSMPlant Situation Model as the literature graph; letting chunks override hard stops; inventing rupees from a sourcebook example figure.
Would change this: Pilot evidence that vector-only (no method ontology) is enough; or that Graph-First is worth the latency for method selection.
Place in the stack#
How to read it.
- OE corpus, playbooks, and case library feed only the advisory partition — never measured rows.
- Dual-family drafting may cite advisory row ids; uncited claims drop.
- Hard gates, constraint evaluator, and calculator path are unchanged.
Build now: Tier A ingest + vector index. Later: Parallel Hybrid default; Ask Adaptive Router (section 10 scope cut).
View Mermaid source
flowchart TB
%% house-style: oe-corpus-advisory
subgraph src["Advisory sources"]
direction LR
corp["OE corpus<br/>vector + sparse graph"]
play["Playbook bullets<br/>lockfile"]
case["Case library"]
end
subgraph part["Context partitions"]
direction LR
led["Advisory partition"]
fam["Dual-family draft"]
end
gate{{"Hard gates unchanged"}}
corp & play & case --> led --> fam --> gate
classDef govc fill:#fff4d6,stroke:#c99a2e,color:#000
classDef agentc fill:#e8f0ff,stroke:#5b7bd5,color:#000
classDef loopc fill:#eef7ee,stroke:#4f9a4f,color:#000
class gate govc
Ingest (summary)#
Detail and license notes: research 20.
| Tier | Examples | v1? |
|---|---|---|
| A — public | DOE sourcebooks (air, pumps, fans, motors, steam, process heat); DOE tip sheets; EPA Lean & Energy toolkit; public IPMVP / ISO 50001 summaries | Yes |
| B — licensed / curated | Factory Physics, TOC, Roser papers → curated method cards; full text only with rights | Selective |
| C — plant-private | SOPs, treasure hunts, OEM manuals — owner-gated, plant-partitioned | Later |
| Out | Marketing % claims; OT-write instructions; chat logs; raw series | Never |
Runtime use#
- Code classifies situation →
ScenarioClass(from family / pattern / scanner). - Allowlisted mission (
oe_methods_for_scenario, …) retrieves top-K chunks + method nodes. - Rows appended to advisory partition with
provenance=doc_id#chunkand corpus version. - Models may cite those row ids. Uncited claims drop.
- Hard gates, constraint evaluator, calculator path unchanged.
Seam id (catalog add): oe_retrieval_mission — options from mission registry; disagreement → registry default (routing class). Jev may later replace mission choice, not the store.
Change class#
| Surface | Class |
|---|---|
| New public PDF in Tier A | Data (ingest job + lockfile corpus pin) |
| New method node / edge | Plug-in / registry |
| New retrieval mission | Registry + replay |
| Letting corpus set money or hard gates | Forbidden (kernel) |
v1 slice vs later#
| v1 | Later |
|---|---|
| Tier A ingest; vector index; hand-curated sparse ontology for Pilot scenarios | Broader Tier B/C; Parallel Hybrid default; Ask Adaptive Router |
| Investigative lane + Ask retrieve OE knowledge | Certified fast path may pull a tiny pinned method card only |
| Offline council proposes playbook bullets from corpus + cases | Same |
| Agentic GraphRAG | Offline curation only |
Links#
- Research deep-dive:
20-oe-knowledge-retrieval.md - Context partitions:
05-context-engineering.md - Memory tiers:
06-memory.md - Improvement:
13-improvement.md
Page history: last 2 changes
- docs(technical): rewrite l4 21-30, glossary and README; reconcile architecture gaps
e7fead7 - docs(l4): agentic decision architecture, ADRs, and production hardness
8275e7c