Industrial AI Agent is an independently developed, locally reproducible full-stack project for examining traceable AI-assisted manufacturing workflows. It uses fictional equipment, synthetic production records, and independently written documents. It does not contain production data, proprietary system material, or a connection to live equipment.
The primary workflow combines one deterministic manufacturing evidence path with Document Search in the same exchange. Manufacturing evidence runs first; only allowlisted recorded fields may enrich retrieval. The interface keeps calculated values, retrieved sources, model interpretation, and missing evidence visibly separate. Co-occurrence is never presented as proof of cause.
v1.0 status: Implemented. The release boundary passed local deterministic verification, browser and screenshot review, public-copy and publication review, two-axis code review, and GitHub Actions on 2026-08-15.
v2.0 status: Implemented. Completed assistant messages now retain typed Evidence Snapshots across reloads. Supported provider working notes remain a temporary, optional UI surface and never become message history or evidence.
- FastAPI and SQLite conversation persistence with explicit Alembic migrations;
- synchronous and SSE assistant exchanges with completed-response persistence and canonical Evidence Snapshots attached to assistant messages;
- temporary SSE Model Working Notes for supported final-answer reasoning, kept separate from assistant messages and Evidence Snapshots;
- deterministic production summary, recorded equipment status, and defect distribution tools over one fictional AOI dataset;
- structured English and Traditional Chinese routing with bounded retry, clarification, and fallback behavior;
- local Markdown retrieval with stable citations, deterministic feature-hashing embeddings, and no external vector database;
- one bounded manufacturing-then-document Combined Evidence workflow;
- a responsive React, Ant Design 6, and Ant Design X workbench;
- an independent local stdio MCP server for three deterministic manufacturing tools; and
- a 45-scenario deterministic offline evaluation suite.
The desktop and 390 px screenshots below show the same accepted local
happy-path observation with qwen3:14b. Model wording is not a reproducibility
target; the route, typed evidence, citations, provenance, and grounding rules
are. See the demo contract for the exact prompt and acceptance
invariants.
The browser sends message requests through the FastAPI SSE boundary. The application owns routing and calls deterministic domain tools before asking an optional OpenAI-compatible model to synthesize an answer. SQLite stores conversations, messages, workspace context, and canonical completed Evidence Snapshots on assistant messages. Current Evidence remains scoped to the active request; reloading history returns each completed snapshot with the assistant message that produced it.
When the configured provider supplies supported final-answer reasoning, the SSE exchange can also show temporary Model Working Notes. They remain separate from the answer and deterministic evidence, are not persisted with message history, and do not expose routing or tool-selection reasoning.
See Architecture for the system boundary and Combined Evidence sequence.
Prerequisites are Python 3.12, uv, Node.js 24,
and npm. The API can start without a model, but assistant requests require an
OpenAI-compatible service and LLM_MODEL.
cd apps/api
uv sync --locked
uv run alembic upgrade head
uv run uvicorn industrial_agent.main:app --host 127.0.0.1 --port 8000In another terminal:
cd apps/web
npm ci
npm run devVite proxies /api to http://127.0.0.1:8000. Configuration and endpoint
details are in the API guide and
Web guide.
cd apps/api
uv sync --locked
uv run alembic upgrade head
uv run pytest -q
uv run ruff check .
uv build
uv run industrial-agent-eval
cd ../web
npm ci
npm test -- --run
npm run typecheck
npm run lint
npm run buildThe latest recorded results and release evidence are listed in Implementation Status. GitHub Actions runs the same deterministic boundary on Ubuntu. Model calls and secrets are excluded from CI.
The final v2.0 local verification recorded 466 full API tests and 145 full Web
tests. Deterministic evaluation, migration, type checking, linting, and builds
passed; existing Fast Refresh and chunk warnings remain.
Ant Design CLI info, lint, and doctor checks remain blocked by the missing
@oxc-parser/binding-darwin-arm64 package. A local browser acceptance run also
passed with an independently created deterministic OpenAI-compatible streaming
fixture. Historical production, Combined partial-failure, deleted-upload source,
missing, and unsupported-version states survived reload. Working Notes opened,
collapsed at the Final Answer, reopened, showed truncation and interruption,
and disappeared on reload. At 390 px, the page had no horizontal overflow.
- Numeric manufacturing results come from deterministic Python code.
- Equipment status comes from explicit synthetic intervals, not inference.
- Document citations refer only to repository-owned fictional Markdown or a user's local upload.
- Model text is interpretation. It is not an equipment command, verified root cause, or production decision.
- A failed assistant request keeps the user message. A real-socket integration test verifies that disconnecting before completion does not persist a partial assistant message; upstream provider cancellation is not guaranteed.
This is a local, single-user application. It has no authentication, multi-tenant isolation, public deployment stack, live equipment integration, persistent vector store, PDF/OCR ingestion, an evidence browser or complete evidence timeline, pagination, persistent Model Working Notes history, full trace or replay, a ThoughtChain surface, or arbitrary planner-driven multi-tool execution. Local Markdown uploads may be sent to the configured model service when retrieved.
Dependency findings and their current exposure are recorded in the Security Review. Evaluation scope and raw-artifact handling are documented in Evaluation.
- Documentation index
- Project scope
- Architecture
- Demo
- Evaluation
- Security review
- Implementation status
- Roadmap
This project is available under the MIT License.