feat: Implement TalkHier framework for hierarchical multi-agent content generation

[harshalmore31]

This PR introduces the TalkHier framework, a novel hierarchical multi-agent system designed for structured content generation and refinement. TalkHier leverages a team of specialized agents, each with a distinct role, to collaboratively produce high-quality content through iterative communication and evaluation.

Key Features:

• Hierarchical Multi-Agent System: TalkHier employs a hierarchical structure with agents specializing in different aspects of content creation:
  • Supervisor: Orchestrates the entire process, analyzes tasks, selects evaluation criteria, and guides revision.
  • Generator: Creates initial content based on instructions.
  • Evaluator (Multiple): Critically assesses content quality against dynamically generated criteria.
  • Revisor: Refines content based on evaluator feedback.
  • Criteria Generator: Dynamically generates task-specific evaluation criteria.
• Structured Communication: Agents communicate using structured CommunicationEvent objects, ensuring clear and organized information exchange in JSON format. This includes messages, background context, intermediate outputs, sender/receiver information, and timestamps.
• Iterative Refinement: The system iteratively refines content through evaluation and revision cycles, aiming to meet a defined quality threshold or a maximum number of iterations.
• Dynamic Evaluation Criteria: TalkHier dynamically generates evaluation criteria tailored to each specific content generation task, ensuring relevant and targeted feedback.
• Parallel Evaluation: Utilizes ThreadPoolExecutor to enable parallel content evaluation by multiple evaluator agents, speeding up the refinement process.
• Robust JSON Handling: Implements robust JSON parsing with error handling and fallback mechanisms to ensure smooth communication between agents, even with variations in LLM output formats.
• State Management: Includes functionality for saving and loading agent states, allowing for persistence and continuation of content generation processes.
• Configurable Parameters: Offers configurable parameters such as max_iterations, quality_threshold, and model_name to customize the content generation process.
• Conversation Tracking: Maintains a Conversation object to log agent interactions and the overall content generation history for debugging and analysis.

Benefits:

• Improved Content Quality: The hierarchical and iterative approach, combined with structured evaluation and revision, leads to higher quality content compared to single-agent systems.
• Task-Specific Content: Dynamic criteria generation ensures content is evaluated and refined based on the specific requirements of each task.
• Enhanced Collaboration: The framework demonstrates effective collaboration between specialized agents, mimicking a team-based content creation workflow.
• Transparency and Explainability: Structured communication and conversation logging provide insights into the content generation process and agent decision-making.
• Extensibility: The modular design allows for easy extension and customization, such as adding new agent roles or modifying existing agent behaviors.

---

📚 Documentation preview 📚: https://swarms--783.org.readthedocs.build/en/783/

[greptile-apps]

PR Summary

The TalkHier framework introduces a sophisticated hierarchical multi-agent system for content generation and refinement. Here are the key points from analyzing the implementation:

• Implements a robust hierarchical structure in /swarms/structs/Talk_Hier.py with specialized agents (supervisor, generator, evaluator, revisor) communicating through structured CommunicationEvent objects
• Utilizes ThreadPoolExecutor for parallel evaluation processing to improve performance when multiple evaluators assess content simultaneously
• Adds dynamic criteria generation through a dedicated CriteriaGenerator agent that tailors evaluation metrics based on task requirements
• Includes comprehensive error handling and fallback mechanisms for JSON parsing, though the current implementation may mask important errors by converting any text to a dict
• The example code contains a potential typo using model name "gpt-4o" which should be addressed for proper functionality

_{💡 (1/5) You can manually trigger the bot by mentioning @greptileai in a comment!}

_{1 file(s) reviewed, 3 comment(s)
Edit PR Review Bot Settings | Greptile}

[greptile-apps]

syntax: model_name='gpt-4o' appears to be a typo and should be 'gpt-4'

Suggested change

	model_name="gpt-4o",
	model_name="gpt-4",

[greptile-apps]

style: ThreadPoolExecutor should specify max_workers to prevent resource exhaustion with many evaluators

[greptile-apps]

logic: regex pattern r"{.}" is greedy and may match too much - use r"{.?}" for non-greedy matching

Suggested change

	json_match = re.search(r"\{.*\}", json_str, re.DOTALL)
	if (json_match):
	json_match = re.search(r"\{.*?\}", json_str, re.DOTALL)
	if (json_match):