VLMForge v0.1.0 — Private-image VLM selection with LangGraph

VLMForge v0.1.0 is the first public release of a local-first decision workbench for selecting visual models against your own images, JSON Schema, business rules, latency, and cost constraints.

Highlights

• Schema-first private image datasets with bounded JSONLogic business rules.
• Replay, OpenRouter, and OpenAI-compatible provider adapters.
• Deterministic quality, consistency, latency, and cost metrics with bootstrap confidence intervals.
• Hard deployment gates followed by an explainable quality/cost/P95 Pareto decision.
• Durable LangGraph orchestration with Send, PostgreSQL checkpoints, anonymous human-review interrupts, resume, and cancellation.
• FastAPI, PostgreSQL, Redis, ARQ, MinIO, SSE replay, Vue 3 workbench, CLI, and HTML/Markdown/JSON/CSV reports.
• A no-key synthetic retail Replay demo for a complete five-minute evaluation flow.

Quick start

git clone https://github.com/To3akaRin/VLMForge.git
cd VLMForge
cp .env.example .env
docker compose up --build -d

Open http://localhost:4173 and select Start Replay demo.

Verification

• 105 Python tests passed; Ruff and strict Mypy passed across 109 source files.
• 6 Vitest tests and 6 Chromium desktop/mobile Playwright flows passed.
• Docker Compose readiness and the complete Replay/report smoke flow passed with PostgreSQL, Redis, MinIO, API, Worker, and Web.
• The wheel and source distribution include the MIT License; SHA-256 checksums are attached.

Known boundaries

• Real paid OpenRouter/OpenAI-compatible smoke calls are opt-in and were not executed for this release.
• v0.1 is a trusted local single-workspace tool without authentication; do not expose it directly to the public internet.
• Docker Compose is the only supported deployment topology in v0.1.

See the README, methodology, and changelog for details.