How mohit67890/imajev works
Imajev is a local typed-decision model: a caller supplies questions, application state and optional images, and receives calibrated probabilities plus an explicit abstention option. Its playground serves this workflow through an HTTP API; model backends compile and score the questions, while shared contracts and calibration shape the response. ImajevBench provides a separate submission-evaluation and review workflow. The graph also preserves the documented live examples and benchmark capabilities. Connections beyond the sampled excerpts are included only where the README or SOURCE INDEX supports them.
Main entry points: server.py HTTP playground · cli.py Benchmark CLI · contracts.py Request contracts · harness.py Evaluation harness
Components 20
Decision serving
- HTTP playground
scripts/playground/server.py - Request contracts
src/vision_decision/contracts.py - Jev request mapping
src/vision_decision/jev_api.py - Image handling
src/vision_decision/images.py
ImajevBench
- Benchmark CLI
src/imajev_bench/cli.py - Dataset schema
src/imajev_bench/schema.py - Submission runner
src/imajev_bench/runner.py - Evaluation harness
src/imajev_bench/harness.py - Benchmark scoring
src/imajev_bench/scoring.py - Blind review packets
src/imajev_bench/review.py - Dataset linting
src/imajev_bench/lint.py - Kimi review workflow
scripts/p3/kimi_review.py
Model runtime
- MLX inference
src/vision_decision/backend.py - Decision scoring
src/vision_decision/scoring.py - Probability calibration
src/vision_decision/calibration.py - Thinking support
src/vision_decision/thinking.py - Remote thought generation
scripts/thought_engine.py - Torch inference
scripts/torch_decision.py
Demo experience
- Checked examples
scripts/playground/examples.py
Other components
- Application caller
Connections 24
- Application caller → HTTP playground: sends request
README.md - HTTP playground → Request contracts: validates request
- HTTP playground → Jev request mapping: maps request
scripts/playground/server.py - HTTP playground → Image handling: handles images
scripts/playground/server.py - HTTP playground → MLX inference: invokes backend
scripts/playground/server.py - HTTP playground → Probability calibration: applies calibration
scripts/playground/server.py - HTTP playground → Thinking support: configures thinking
scripts/playground/server.py - MLX inference → Decision scoring: compiles questions
src/vision_decision/backend.py - Decision scoring → Request contracts: uses types
src/vision_decision/scoring.py - Probability calibration → Request contracts: returns results
src/vision_decision/calibration.py - Thinking support → Request contracts: uses types
src/vision_decision/thinking.py - Thinking support → Remote thought generation: requests thoughts
- Torch inference → Decision scoring: uses scoring
scripts/torch_decision.py - HTTP playground → Checked examples: loads examples
scripts/playground/server.py - Benchmark CLI → Submission runner: dispatches runs
src/imajev_bench/cli.py - Benchmark CLI → Blind review packets: dispatches review
src/imajev_bench/cli.py - Benchmark CLI → Dataset schema: dispatches checks
src/imajev_bench/cli.py - Benchmark CLI → Benchmark scoring: dispatches scoring
src/imajev_bench/cli.py - Benchmark CLI → Dataset linting: dispatches lint
src/imajev_bench/cli.py - Submission runner → Dataset schema: validates records
src/imajev_bench/runner.py - Evaluation harness → Submission runner: executes submissions
src/imajev_bench/harness.py - Evaluation harness → Dataset schema: uses records
src/imajev_bench/harness.py - Evaluation harness → Benchmark scoring: scores outputs
src/imajev_bench/harness.py - Kimi review workflow → Blind review packets: supports review