How githubnext/localjev works
LocalJev is a Jev-compatible HTTP bridge: clients submit state and typed questions, the server validates them, and the engine prompts an OpenAI-compatible model for probabilities before returning Jev-shaped answers. A separate documented capability runs reproducible model bake-offs and reports quality, calibration, and latency. The evaluation branch is included because its substantial runtime modules are part of the repository; relationships inside the engine workflow follow the README, since the sampled engine excerpt omits those implementation details.
Main entry points: run.ts Bake-off Runner · index.ts Bun Entrypoint · engine.ts Decision Engine · server.ts HTTP Routes
Components 18
Jev API
- Bun Entrypoint
src/index.ts - Runtime Settings
src/config.ts - HTTP Routes
src/server.ts - Request Validation
src/types.ts
Inference
- Decision Engine
src/engine.ts - Classification Prompt
src/engine.ts - Chat Completion
src/engine.ts - Probability Parsing
src/engine.ts - Jev Answers
src/engine.ts
Model Evaluation
- Bake-off Runner
scripts/eval/run.ts - Dataset Preparation
scripts/eval/data.ts - Evaluation Types
scripts/eval/common.ts - Report Generation
scripts/eval/report.ts - Quality Metrics
scripts/eval/metrics.ts
Other components
- API Client
- OpenAI-Compatible Model
- Public Datasets
- Evaluation Run Files
Connections 22
- API Client → Bun Entrypoint: sends request
README.md - Bun Entrypoint → Runtime Settings: loads settings
src/index.ts - Bun Entrypoint → HTTP Routes: serves requests
src/index.ts - HTTP Routes → Request Validation: validates body
src/server.ts - HTTP Routes → Decision Engine: requests decision
src/server.ts - Decision Engine → Runtime Settings: uses settings
src/engine.ts - Decision Engine → Classification Prompt: prepares prompt
README.md - Classification Prompt → Chat Completion: requests probabilities
README.md - Chat Completion → OpenAI-Compatible Model: calls endpoint
src/engine.ts - Chat Completion → Probability Parsing: parses output
README.md - Probability Parsing → Jev Answers: normalizes and scores
README.md - Bun Entrypoint → Bake-off Runner: runs separately
- Bake-off Runner → Runtime Settings: loads settings
scripts/eval/run.ts - Bake-off Runner → Decision Engine: runs decisions
scripts/eval/run.ts - Bake-off Runner → Dataset Preparation: prepares examples
scripts/eval/run.ts - Dataset Preparation → Public Datasets: downloads data
scripts/eval/data.ts - Bake-off Runner → Evaluation Types: uses evaluation types
scripts/eval/run.ts - Dataset Preparation → Evaluation Types: uses evaluation types
scripts/eval/data.ts - Bake-off Runner → Report Generation: generates report
scripts/eval/run.ts - Report Generation → Quality Metrics: calculates metrics
scripts/eval/report.ts - Bake-off Runner → Evaluation Run Files: writes results
scripts/eval/run.ts - Report Generation → Evaluation Run Files: writes summaries
scripts/eval/report.ts