Accelerated Discovery Framework

A human-centric MAS Framework for scientific discovery that empowers researchers while maintaining scientific integrity, transparency, and reproducibility.

Core Philosophy

This framework is built on the principle that human researchers should direct the discovery process. AI agents are powerful tools to augment, not replace, human intellect and intuition. We prioritize:

• Human-in-the-loop control - The researcher has final say on workflow, parameters, and interpretation
• Scientific integrity - Deep attribution, conflicting evidence identification, and rigorous validation
• Transparent & reproducible research - Every workflow is a shareable, inspectable artifact
• Open collaboration - Community-driven framework for shared scientific advancement

Architecture

The system implements a Planner-Orchestrator pattern with standardized NodeTemplate components for maximum flexibility and scientific rigor:

Core Design Patterns

• NodeTemplate Architecture: All functional components implement the standardized AbstractNodeTemplate (akd.nodes.templates) with:

  • Well-defined state management
  • Input/output guardrails for validation
  • Tool subset isolation (principle of least privilege)
  • Framework-agnostic design that can wrap into any orchestration engine

• Context Management:

  • Global Context: Maintains overall research project state
  • Local Context: Sandboxed, need-to-know subsets for individual agents
  • Prevents context bleeding between agents

• Human-in-the-Loop Control: Researchers maintain control over:

  • Workflow approval and modifications
  • Parameter tuning for any component
  • Branching and merging decisions

• Multi-Agent Coordination: Specialized agents work together with conflict detection and gap identification

Human-in-the-Loop Control Points

The framework ensures researchers maintain control throughout the discovery process:

• Plan Approval: Initial research plans and any significant modifications require explicit human approval
• Parameter Control: Researchers can inspect and adjust parameters for any NodeTemplate component
• Workflow Direction: AI proposes next steps but humans direct the overall research strategy
• Quality Gates: Human validation required before accepting AI-generated analyses or conclusions
• Branching Decisions: Research workflow branching and merging decisions are human-directed

Golden Rule: When in doubt about research direction, data validity, or result interpretation, the system defers to human researcher guidance.

For comprehensive design principles, see Design Philosophy.

Quick Start

Prerequisites

• Python 3.12+
• uv package manager (recommended)

Installation

# Create and activate virtual environment
uv venv --python 3.12
source .venv/bin/activate

# Install dependencies
uv sync

# For development (includes testing tools)
uv sync --extra dev

# For local development (includes marimo and other local tools)
uv sync --extra dev --extra local

# For ML development (includes pandas, sentence-transformers, docling, deepeval)
uv sync --extra ml

# Setup environment variables
cp .env.example .env
# Edit .env with your API keys and configurations

Basic Usage

Refer to the notebooks for examples.

Core Tools & Agents

The framework provides a comprehensive suite of specialized tools and agents for scientific research:

Search Tools

• SearxNGSearchTool (akd.tools.search.searxng_search) - General web search with privacy focus
• SemanticScholarSearchTool (akd.tools.search.semantic_scholar_search) - Academic paper search and discovery
• Code Search Tool (akd.tools.code_search) - Code repository search and analysis

Search Agents

• Deep Search Agent (akd.agents.search.deep_search) - Advanced multi-step literature search with iterative refinement
• Controlled Search Agent (akd.agents.search.controlled) - Structured search workflow with quality controls
• Query Agent & FollowUp Query Agent (akd.agents.query) - Query processing, refinement, and follow-up generation
• Relevancy Agent (akd.agents.relevancy) - Content relevance assessment and filtering

Extraction & Processing

• PyPaperBot Scraper (akd.tools.scrapers.pypaperbot) - PDF content extraction and processing
• DoclingScraper (akd.tools.scrapers) - Advanced document processing and text extraction
• Web Scrapers (akd.tools.scrapers.web_scrapers) - Content extraction from web sources

Specialized Components

• Source Validator (akd.tools.source_validator) - Source quality validation and verification
• Relevancy Checker (akd.tools.relevancy) - Content relevance validation against research context
• Link Relevancy Assessor (akd.tools.link_relevancy_assessor) - URL relevance assessment and filtering

Scientific Guardrails

The framework implements deep guardrails specifically designed for scientific research integrity:

Deep Attribution & Validation

• Traceable Claims: All claims traceable to specific source sentences and data points
• Source Quality: Prioritizes refereed journals and validated data repositories
• Attribution Chain: Complete attribution from final claims back to original sources
• Quality Validation: Multi-level validation of source credibility and relevance

Conflict Detection & Gap Analysis

• Agentic RAG: Multi-agent approach to comprehensive information retrieval
  • Gap Agent: Actively identifies missing information and research gaps
  • Conflict Agent: Specifically searches for contradictory evidence and conflicting findings
• Bias Prevention: Deliberately surfaces contradictory evidence to prevent confirmation bias
• Evidence Balance: Ensures both supporting and conflicting evidence is presented

Transparent Research Process

• Stateful Execution: Complete workflow state capture for reproducibility
• Shareable Artifacts: Research graphs that others can inspect, validate, and extend
• Human Validation: All AI-generated content clearly labeled and requires human approval
• Complete Transparency: Every step of the research process is inspectable and documented

Key Features

• Human-in-the-Loop Control: Researchers direct the discovery process with AI augmentation
• Framework Agnostic: Core logic decoupled from orchestration engines for maximum flexibility
• Reproducible Research: Complete workflow capture enables true reproducibility and sharing
• Community-Driven: Open framework designed for collaborative scientific advancement

Project Structure

akd/                    # Core framework
├── agents/            # Specialized research agents
├── nodes/             # NodeTemplate implementations
├── tools/             # Research tools and scrapers
└── configs/           # Configuration management

examples/              # Usage examples
scripts/               # Utility scripts

Contributing

This is an open, community-driven framework. See our design philosophy for development and design guidelines.

License

This project is licensed under the Apache License 2.0 - see the LICENSE file for details.