ResearchLoom (Samhita) Documentation & Wiki

Architecture, Stylus Dynamics, Flow Maps, and Systematic Literature Synthesis

---

Project maintained by sohamsans Hosted on GitHub Pages — Theme by mattgraham

Evidence Matrix & Two-Stage LLM Synthesis

ResearchLoom achieves fast, accurate, and cost-effective research triage through its strict Two-Stage LLM Architecture.

---

🎯 The Two-Stage Evaluation Pipeline

graph LR
    A[Academic Search: OpenAlex / arXiv] --> B[Deduplicated Candidate Pool]
    B --> C[Stage 1: Batch Triage]
    C -->|Score >= 70%| D[Stage 2: Deep Extraction]
    C -->|Score < 70%| E[Archived / Excluded]
    D --> F[Evidence Matrix & Synthesized Review]

Stage 1: Batch Triage

• Input: Truncated title and abstract (max 180 words per paper).
• Execution: Batches of 5 papers processed simultaneously using lightweight high-speed models (gemini-2.0-flash or Groq llama-3.1-8b-instant).
• Output: Strict JSON schema scoring relevance (0–100), core mechanism, and primary application.

Stage 2: Deep Evidence Extraction

• Input: Targeted slices from the paper’s Open Access PDF (Introduction, Methodology, and Conclusion, limited to 8,000 characters).
• Concurrency: 6 parallel workers streaming extractions simultaneously.
• Output: Quantitative metrics, statistical outcomes, sample sizes, and verbatim grounding quotes.

---

🔑 Multi-Provider BYOK Configuration

You can bring your own API keys for:

• Google Gemini: gemini-2.0-flash, gemini-1.5-flash
• Groq Cloud: llama-3.1-8b-instant (ultra-low latency)
• Anthropic Claude: claude-3-5-sonnet, claude-3-5-haiku
• OpenAI: gpt-4o, gpt-4o-mini
• OpenRouter Free Tier: DeepSeek, Mistral, and Llama community endpoints