An evidence-first, offline prototype for auditable Vietnamese public-service dossier checking.
- Live demo: https://mrlouisdev.github.io/govflow-vn/demo/
- Architecture: docs/architecture.md
- Evaluation protocol: docs/evaluation.md
- Roadmap and maintenance: docs/maintenance.md
GovFlow VN is currently a deterministic offline MVP, not a production administrative system. It demonstrates a narrow, inspectable workflow:
- ingest a structured synthetic dossier;
- normalize its fields;
- detect missing or inconsistent information with deterministic rules;
- attach a source identifier to every finding;
- refuse unsupported questions instead of guessing; and
- return an audit record for evaluation.
OCR ingestion, Vietnamese SLM/RAG, authoritative procedure rulesets, and production deployment are roadmap items. They are not claimed as implemented features.
The independent VietDoc Edge team developed this public MVP while preparing its GovFlow VN concept for A.I Thực Chiến 2026. The repository makes the implemented baseline, synthetic examples, architecture, and evaluation method independently inspectable. This statement records project context only; it does not claim competition selection, endorsement, or an award.
The longer-term research question is whether a small, local-first Vietnamese workflow can provide useful dossier checks while keeping citations, refusal behavior, and resource use measurable.
- Offline deterministic dossier-validation engine.
- Findings with source identifiers for each triggered rule.
- Out-of-scope question guard.
- Reproducible synthetic dossier samples.
- Versioned synthetic benchmark dataset, harness, and machine-readable result.
- Browser demo and automated Node.js tests.
- Architecture and evaluation specifications for future OCR and Vietnamese SLM/RAG adapters.
Prerequisite: Node.js 20 or newer. The current MVP has no runtime package dependencies.
git clone https://github.com/mrlouisdev/govflow-vn.git
cd govflow-vn
node --version
npm test
npm run benchmarknpm test must exit with code 0. npm run benchmark writes the versioned result to benchmarks/results/v0.1.0.json; interpret it only as evidence for the bundled synthetic cases. To inspect the browser demo locally:
npm run serveThen open http://127.0.0.1:8080/demo/. Stop the server with Ctrl+C.
demo/ Browser UI
examples/ Synthetic dossier samples
src/ Deterministic evaluation engine
tests/ Automated tests
benchmarks/ Versioned synthetic evaluation data, harness, and results
docs/ Architecture, evaluation, and maintenance notes
scripts/ Local run helpers
The published evaluation protocol defines missing-field recall, citation coverage and correctness, out-of-scope recall, latency, and memory measurements. Run npm run benchmark to regenerate the bundled synthetic evaluation result. A score should not be reported without its dataset version, engine version, and raw evaluation log.
- Bundled dossiers and rules are synthetic and do not represent an official administrative procedure.
- The MVP accepts structured JSON; it does not currently extract content from PDF files or images.
- Source identifiers refer to the synthetic demo ruleset, not Vietnamese legal authority.
- The scope guard is deterministic and narrow; it is not a general hallucination detector.
- No production privacy, identity, authorization, accessibility, or availability review has been completed.
- Bundled benchmark results describe synthetic cases only; they are not evidence of real-world accuracy, production performance, or adoption.
The primary maintainer is @mrlouisdev. Maintainer scope, review expectations, and the public roadmap are documented in MAINTAINERS.md and docs/maintenance.md.
Contributions are welcome. Read CONTRIBUTING.md before opening an issue or pull request. For security-sensitive reports, follow SECURITY.md. Changes are recorded in CHANGELOG.md.
Licensed under the Apache License 2.0.