title	Open Research Lite
emoji	⚡
colorFrom	indigo
colorTo	blue
sdk	gradio
sdk_version	4.44.0
app_file	app.py
pinned	false
license	mit
short_description	Slash Deep Research agent token bloat by 56%
tags	deep-research llm-agents knowledge-graph agentic-rag token-optimization paper:10.5281/zenodo.22168098

open-research-lite ⚡

Differential Knowledge-State Tracking & Concept-Diff Ingestion Middleware for Autonomous Deep Research Agents

📄 Official Research Paper: Differential Knowledge-State Tracking for Token-Efficient Autonomous Deep Research Agents (Zenodo / CERN)

---

🥊 Head-to-Head Benchmark: Open-Research-Lite vs. GPT-Researcher

In existing research agents (e.g. GPT-Researcher, AutoGPT, standard agentic RAG), over 65% of prompt tokens consist of redundant introductory boilerplate, duplicated corporate bios, and SEO fluff scraped across consecutive queries.

open-research-lite replaces raw document concatenation with dynamic knowledge-state tracking, passing only novel differential fact assertions ($\Delta G_t$) while deterministically isolating metric contradictions.

Research Benchmark Domain	🔴 GPT-Researcher Tokens	🟢 Open-Research-Lite Tokens	⚡ Token & Cost Savings	🎯 Metric Contradictions Isolated
Solid-State EV Batteries (2026)	871 tokens	389 tokens	55.3% Cheaper	2 Flagged (GPT-Researcher: 0)
Quantum QEC Scaling	782 tokens	297 tokens	62.0% Cheaper	0 Clean
HBM4 Memory Interconnect & Power	744 tokens	371 tokens	50.1% Cheaper	2 Flagged (GPT-Researcher: 0)
De Novo Protein Design	555 tokens	311 tokens	44.0% Cheaper	2 Flagged (GPT-Researcher: 0)
HTS Tokamak Magnetic Fusion	679 tokens	336 tokens	50.5% Cheaper	2 Flagged (GPT-Researcher: 0)
TOTAL MULTI-DOMAIN	3,631 tokens	1,704 tokens	53.1% FEWER TOKENS	8 Conflicts Caught

---

🚀 Key Features

• Continuous Knowledge-State Tracking ($G_t = G_{t-1} \cup \Delta G_t$): Maintains an active in-memory session graph of verified factual assertions across multi-turn search loops.
• Deterministic Contradiction Detection: If Source A claims $110/kWh and Source B claims $140/kWh, open-research-lite flags the explicit dispute instead of letting the synthesizer LLM silently average or hallucinate.
• Dual-Layer Extraction Architecture:
  • Layer 1 (LLM Mode): Structured atomic triplet extraction (Subject ──► Predicate ──► Object) via Gemini 2.5 Flash or OpenAI.
  • Layer 2 (Local NLP Mode): Sub-millisecond deterministic regex and grammar extraction running locally for $0 cost.
• Drop-in Middleware: Integrates directly into LangGraph, AutoGPT, CrewAI, or GPT-Researcher in 3 lines of Python.

---

⚡ Quickstart

1. Installation

pip install open-research-lite

2. Basic Usage (Python API)

import asyncio
from open_research_lite import ConceptDiffEngine

async def main():
    # Automatically extracts novel differential facts from scraped text
    engine = ConceptDiffEngine()

    raw_scraped_text = """
    Electric vehicles have become popular over the last decade... 
    Lithium-ion batteries were invented by John Goodenough...
    In 2026, researchers demonstrated a solid-state cell achieving 500 Wh/kg energy density.
    Vendor targets pilot production cell cost at $110/kWh.
    """

    # Process raw scrape into a condensed Diff Payload
    diff_payload = await engine.process_observation(
        raw_text=raw_scraped_text,
        source_url="https://autonews.com/battery-2026",
        source_title="2026 Battery Report"
    )

    print("--- HIGH-SIGNAL DIFF PAYLOAD ---")
    print(diff_payload)

if __name__ == "__main__":
    asyncio.run(main())

---

🔬 Reproduce the Benchmarks Locally

You can run the full multi-domain benchmark evaluation suite with:

git clone https://github.com/vishal-raaj-dnd/open-research-lite
cd open-research-lite
pip install -e .
python benchmark_vs_gpt_researcher.py

---

📜 Citation

If you use open-research-lite or the Concept-Diff framework in your research, please cite our official paper:

@article{raaj2026conceptdiff,
  title={Differential Knowledge-State Tracking for Token-Efficient Autonomous Deep Research Agents},
  author={Raaj, Vishal},
  journal={Zenodo Preprint},
  year={2026},
  doi={10.5281/zenodo.22168098},
  url={https://doi.org/10.5281/zenodo.22168098}
}

---

📄 License

MIT License. Open-sourced by the Open-Research-Lite Initiative.