RFC: ManufacturingAgentSuite blueprint for route-isolated industrial action-card agents

[Ryanhuii]

Summary

I would like to propose a new OPEA GenAIExamples blueprint named ManufacturingAgentSuite.

The blueprint demonstrates a route-isolated manufacturing agent suite that uses one OPEA-style Gateway and Megaservice to coordinate route-specific Dataprep, Retriever/RAG, Vector DB, LLM, deterministic evaluation, guardrails, and action-card outputs.

The first reference implementation is WearEdge OPEA Manufacturing:

https://github.com/davidmillerak2026-sys/wearedge-opea-manufacturing

Why This Belongs In OPEA

Manufacturing deployments often need more than a single chatbot. A plant-floor assistant must route the same evidence stream to different governed workflows:

Route	Manufacturing workflow	Output target
maintenance	Predictive maintenance / lao-shi-fu troubleshooting	CMMS work-order draft
iqc	Incoming and in-process quality control	QMS quality event
changeover	SKU changeover verification	MES/checklist hold
wi	Released work-instruction guidance	Work-instruction reference
hazard	EHS hazard observation	EHS observation/case

This pattern can help OPEA users understand how to build enterprise agent suites where every route has its own retrieval source, output contract, guardrails, and integration target.

Proposed Blueprint Shape

ManufacturingAgentSuite/
  README.md
  manufacturing_agent_suite.py
  docker_compose/intel/cpu/xeon/compose.yaml
  benchmark/README.md
  tests/test_compose_on_xeon.sh
  assets/flow.md

The blueprint should show:

• Gateway endpoint for /v1/agents.
• Route-specific /v1/agents/{mode}/demo and /v1/agents/{mode}/infer.
• Route registry with fixed input examples and output action-card contracts.
• Dataprep + Retriever/RAG + Vector DB collections isolated by route.
• Guardrails blocking unsupported release, restart, root-cause, and safety-clearance claims.
• A scorecard endpoint that reports latency, contract pass, guardrail pass, RAG/source match, action target correctness, and route isolation.

OPEA Component Mapping

The reference implementation maps to existing OPEA concepts:

OPEA concept	Blueprint use
Gateway	Plant evidence/API entry point
Megaservice	Route orchestration and service composition
Dataprep	Route-specific manuals, quality plans, policies, and checklists
Retriever/RAG	Source-grounded action-card generation
Vector DB	Qdrant route collections, with local fallback in the reference package
LLM service	Pluggable LLM adapter, deterministic no-model path for CI
Guardrails	Blocked claims and human confirmation gates
Evaluation	Deterministic route scorecard

Reference Validation

The reference repository currently supports:

docker compose up --build -d
curl http://127.0.0.1:8088/v1/agents
curl http://127.0.0.1:8088/v1/agents/maintenance/demo
curl http://127.0.0.1:8088/v1/agents/iqc/demo
curl http://127.0.0.1:8088/v1/agents/changeover/demo
curl http://127.0.0.1:8088/v1/agents/wi/demo
curl http://127.0.0.1:8088/v1/agents/hazard/demo
curl http://127.0.0.1:8088/v1/scorecard

Expected route registry:

maintenance, iqc, changeover, wi, hazard

Requested Feedback

I would appreciate maintainer feedback on:

• Whether this should be contributed as a new GenAIExamples/ManufacturingAgentSuite example.
• Whether route-isolated guardrails and action-card contracts belong in GenAIExamples or GenAIComps.
• Whether OPEA would prefer a minimal Docker Compose-only blueprint first, with Helm/GMC support added in a follow-up PR.
• Naming guidance for the example and route registry endpoints.

Follow-Up PR Plan

If maintainers agree, I can prepare a PR with:

• Minimal Docker Compose deployment.
• README and architecture diagram.
• Five sample manufacturing routes.
• Deterministic CI-friendly demo mode.
• Qdrant vector DB profile.
• Scorecard endpoint and E2E test.

[Ryanhuii]

Implementation update for the proposed ManufacturingAgentSuite blueprint:

The reference package is now more judge- and maintainer-friendly. It includes a browser demo console in addition to the API endpoints, so reviewers can inspect route selection, sample requests, RAG evidence, bounded action cards, blocked claims, and the five-route scorecard without writing curl commands.

Reference implementation:
https://github.com/davidmillerak2026-sys/wearedge-opea-manufacturing

Runnable entry point after docker compose up --build -d:

http://127.0.0.1:8088/demo

Current validation status:

• GET /v1/agents returns maintenance, iqc, changeover, wi, and hazard.
• GET /v1/agents/{mode}/demo returns a route-specific action card for every mode.
• GET /v1/scorecard reports five passing route checks covering contract, guardrail, RAG/source match, action target correctness, and route isolation.
• The Docker profile starts a Qdrant-backed gateway, with a deterministic no-model demo path for CI/reviewer reproducibility.

PR-ready minimal scope I can prepare if this direction is useful:

ManufacturingAgentSuite/
  README.md
  manufacturing_agent_suite.py
  docker_compose/intel/cpu/xeon/compose.yaml
  benchmark/README.md
  tests/test_compose_on_xeon.sh
  assets/flow.md

A short technical article describing the route-isolated manufacturing action-card pattern is also public here:
https://github.com/davidmillerak2026-sys/wearedge-opea-manufacturing/blob/main/public/article-wear-edge-opea-manufacturing.md

The main feedback I am looking for before opening the first PR is whether OPEA would prefer this as a new GenAIExamples/ManufacturingAgentSuite example, a docs-first blueprint proposal, or a split where reusable action-card/route-registry pieces move toward GenAIComps later.

[Ryanhuii]

Implementation update: the reference package now includes an official OPEA TEI
embedding profile in addition to the deterministic and OPEA-compatible profiles.

New profile:

docker compose -f docker-compose.yml -f docker-compose.opea-tei.yml up -d

What changed:

• docker-compose.opea-tei.yml wires Hugging Face TEI, opea/embedding:latest,
  Qdrant, and the Manufacturing Gateway.
• The profile uses the OPEA TEI pattern with TEI_EMBEDDING_ENDPOINT,
  EMBEDDING_COMPONENT_NAME=OPEA_TEI_EMBEDDING, and /v1/embeddings.
• Local E2E passed with BAAI/bge-base-en-v1.5, 768-dimensional embeddings,
  all five manufacturing demo routes, and /v1/scorecard passing.
• Google Cloud C3 fresh-clone E2E passed on c3-standard-4 in
  us-central1-a; the temporary VM wearedge-opea-tei-0527103938 was deleted
  after the run.
• All five route demos report qdrant-opea-tei-vector-store in the RAG
  evidence when this profile is used.
• A Google Cloud C3 fresh-clone rerun script is included for the same profile:
  scripts/gcp_c3_opea_tei_profile_e2e_cloudshell.sh.

Evidence:

• Official TEI profile docs:
  https://github.com/davidmillerak2026-sys/wearedge-opea-manufacturing/blob/main/docs/official-opea-tei-profile.md
• Local TEI E2E report:
  https://github.com/davidmillerak2026-sys/wearedge-opea-manufacturing/blob/main/docs/local-opea-tei-profile-e2e-report.md
• Local TEI JSON summary:
  https://github.com/davidmillerak2026-sys/wearedge-opea-manufacturing/blob/main/evidence/benchmarks/local_opea_tei_profile_e2e.summary.json
• GCP C3 TEI E2E report:
  https://github.com/davidmillerak2026-sys/wearedge-opea-manufacturing/blob/main/docs/gcp-c3-opea-tei-profile-e2e-report.md
• GCP C3 TEI JSON summary:
  https://github.com/davidmillerak2026-sys/wearedge-opea-manufacturing/blob/main/evidence/benchmarks/gcp_c3_opea_tei_profile_e2e.summary.json
• Commit:
  davidmillerak2026-sys/wearedge-opea-manufacturing@e7126c6

This makes the proposed ManufacturingAgentSuite blueprint closer to an
OPEA-native example rather than only an API-compatible wrapper: it now has
Gateway/Megaservice, route-isolated RAG/Qdrant, guardrails, scorecard, and an
official TEI embedding component path.