ai-sci-agent 0.8.0

Full-process autonomous scientific research agent — from literature search to LaTeX paper

🧪 AI-Sci: Autonomous Scientific Research Agent

One command from research question to LaTeX paper. Built on DeepAgents + LangGraph, incorporating the ARIS research methodology.

Install

uv tool install ai-sci-agent
ai-sci setup          # interactive wizard — pick provider, enter key
ai-sci run "improve LoRA with adaptive rank" --paper paper.pdf --repo ~/code

Or with pip:

pip install ai-sci
ai-sci setup

9-Phase Research Pipeline

search → analyze → refine → experiment_plan → reproduce → implement → experiment → review → write

Phase	What it does
search	Multi-source literature search with 3-layer anti-hallucination verification
analyze	PICO-M deep reading, 4-axis critical scoring, cross-paper synthesis
refine	Problem Anchor freeze, 7-dim iterative method scoring
experiment_plan	Claim-driven experiment design with 5-block evidence storyline
reproduce	Clone repos, install deps, run baselines, compare with paper claims
implement	Map method proposal to code changes, edit codebase, verify
experiment	Code integrity audit → sanity check → execute → stats → iterate
review	Structured peer review with weakness grading
write	LaTeX paper + 5-pass sciwrite audit + verified citations

Usage

# Full pipeline
ai-sci run "research question" \
  --paper https://arxiv.org/abs/2106.09685 \
  --repo ~/code \
  --data ~/datasets

# Interactive mode
ai-sci run "explore and improve this codebase" \
  --repo ~/code --interactive

# List/resume sessions
ai-sci sessions
ai-sci run "..." --resume sci-20260623-094430

# Single-phase commands
ai-sci search "query"
ai-sci analyze "Paper Title"

Configuration

Run ai-sci setup for interactive setup, or create ~/.ai_sci/.env:

MODEL_NAME=deepseek-chat
API_KEY=sk-your-key
BASE_URL=https://api.deepseek.com/v1
TAVILY_API_KEY=           # optional

Supports any OpenAI-compatible API: DeepSeek, OpenAI, SiliconFlow, Ollama, vLLM, Groq.

Key Features

• Autonomous end-to-end: one command from question to paper
• Honest: no repo → skips experiments with [NOT VALIDATED]. No data → clearly reported. No GPU → tagged.
• Persistent: SqliteSaver + RESEARCH_WIKI.md + sessions + memory across restarts
• ARIS methodology: Problem Anchor, 7-dim scoring, claim-driven, code-review-first
• Anti-hallucination: 3-layer paper verification, never fabricates citations or results

License

MIT