How varsha-1428/multi-agent-system works
This repository generates three candidate images from a user prompt, scores prompt alignment and cross-image consensus, and iteratively improves only failed candidates using vision feedback. LangGraph coordinates generation, evaluation, pass/fail routing, and regeneration; evaluation combines CLIP-based measures into a thresholded result. The README documents weighted scoring and the feedback loop. The scorer implementation was not excerpted, so its role and connection are supported by the README and source index rather than inspected source.
Main entry points: image_generation_graph.py Workflow Graph · agent1.py Agent 1 · agent2.py Agent 2 · agent3.py Agent 3
Components 16
Workflow Orchestration
- Application Entry
main.py - Workflow Graph
graph/image_generation_graph.py - Workflow State
workflow/state.py
Image Generation
- Agent 1
agents/agent1.py - Agent 2
agents/agent2.py - Agent 3
agents/agent3.py - Agent Interface
agents/base_agent.py - Generated Images
images
Image Evaluation
- Description Matcher
evaluator/description_match.py - Consensus Validator
evaluator/consensus.py - Evaluation Coordinator
evaluator/evaluate.py - Score Combiner
evaluator/scorer.py
Feedback Refinement
- Vision Feedback
evaluator/feedback.py
Other components
- User Prompt
- Hugging Face Inference
- Groq Vision Model
Connections 24
- User Prompt → Application Entry: provides prompt
main.py - Application Entry → Workflow Graph: invokes graph
main.py - Workflow Graph → Workflow State: uses schema
graph/image_generation_graph.py - Workflow Graph → Agent 1: generates with
graph/image_generation_graph.py - Workflow Graph → Agent 2: generates with
graph/image_generation_graph.py - Workflow Graph → Agent 3: generates with
graph/image_generation_graph.py - Agent 1 → Agent Interface: implements
agents/agent1.py - Agent 2 → Agent Interface: implements
agents/agent2.py - Agent 3 → Agent Interface: implements
agents/agent3.py - Agent 1 → Hugging Face Inference: requests image
agents/agent1.py - Agent 2 → Hugging Face Inference: requests image
agents/agent2.py - Agent 3 → Hugging Face Inference: requests image
agents/agent3.py - Agent 1 → Generated Images: saves image
agents/agent1.py - Agent 2 → Generated Images: saves image
agents/agent2.py - Agent 3 → Generated Images: saves image
agents/agent3.py - Workflow Graph → Evaluation Coordinator: evaluates images
graph/image_generation_graph.py - Evaluation Coordinator → Description Matcher: scores alignment
evaluator/evaluate.py - Evaluation Coordinator → Consensus Validator: scores consensus
evaluator/evaluate.py - Evaluation Coordinator → Score Combiner: combines scores
evaluator/evaluate.py - Workflow Graph → Vision Feedback: requests feedback
graph/image_generation_graph.py - Vision Feedback → Groq Vision Model: analyzes image
evaluator/feedback.py - Workflow Graph → Agent 1: regenerates failed
graph/image_generation_graph.py - Workflow Graph → Agent 2: regenerates failed
graph/image_generation_graph.py - Workflow Graph → Agent 3: regenerates failed
graph/image_generation_graph.py