Screen research papers against a research question using six scoring dimensions. Choose criteria for a broad knowledge collection or a systematic review, then receive inclusion, exclusion, and human-review lists with evidence quotes checked against abstracts.
中文简介正在准备中。当前显示英语简介。
Screening Assistant - AI-PRISMA 6-dimension screening with Groq LLM (100x cheaper) Supports two project types with different confidence thresholds Use when: screening papers, PRISMA screening, inclusion/exclusion criteria Triggers: screen papers, PRISMA screening, inclusion criteria, exclusion criteria, AI screening
实测评分 未实测
查看安装方法安装前可以查看交给 AI 代理的指令内容。
diverga_check_prerequisites("i2") → must return approved: true
If not approved → AskUserQuestion for each missing checkpoint (see .claude/references/checkpoint-templates.md)
diverga_mark_checkpoint("SCH_SCREENING_CRITERIA", decision, rationale)Read .research/decision-log.yaml directly to verify prerequisites. Conversation history is last resort.
Agent ID: I2 Category: I - Systematic Review Automation Tier: MEDIUM (Sonnet) Icon: 📋✅
Executes AI-assisted PRISMA 2020 screening using a 6-dimension rubric. Leverages Groq LLM for 100x cost reduction compared to Claude, while maintaining screening quality. Supports two project types with different confidence thresholds.
| Provider | Model | Cost per 100 papers | Quality |
|---|---|---|---|
| Groq (Default) | llama-3.3-70b | $0.01 | Excellent |
| Groq | qwen-qwq-32b | $0.008 | Good |
| Claude | claude-haiku-4-5 | $0.15 | Excellent |
| Claude | claude-sonnet-3-5 | $0.45 | Best |
| Ollama | llama3.2:70b | $0 | Good (local) |
Recommendation: Use Groq for screening. Switch to Claude only for complex edge cases.
Required:
- project_path: "string"
- research_question: "string"
- project_type: "enum[knowledge_repository, systematic_review]"
Optional:
- llm_provider: "enum[groq, claude, ollama]"
- custom_criteria: "object"
- max_workers: "int"
- batch_size: "int"
main_output:
stage: "prisma_screening"
project_type: "string"
threshold: "int"
llm_provider: "string"
model: "string"
results:
total_screened: "int"
auto_included: "int"
auto_excluded: "int"
human_review: "int"
cost:
input_tokens: "int"
output_tokens: "int"
total_cost: "string"
output_files:
relevant_papers: "string"
excluded_papers: "string"
human_review: "string"
Before executing screening, I2 MUST:
PRESENT screening criteria:
AI-PRISMA 6-Dimension Screening Criteria
Project Type: {knowledge_repository | systematic_review}
Threshold: {50% | 90%} confidence
Scoring Rubric:
1. DOMAIN (0-10): Target population/context relevance
2. INTERVENTION (0-10): Technology/tool focus
3. METHOD (0-5): Study design rigor
4. OUTCOMES (0-10): Measured results clarity
5. EXCLUSION (-20 to 0): Penalties for wrong domain/review
6. TITLE BONUS (0 or 10): Keywords in title
Total Score Range: -20 to 50 points
Decision Rules:
- score ≥ {threshold} → auto-include
- score < 0 → auto-exclude
- otherwise → human-review
Do you approve these criteria?
WAIT for explicit approval
CONFIRM before executing screening
# Project path (set to your working directory)
cd "$(pwd)"
# Set LLM provider (v1.2.6: Groq default)
export LLM_PROVIDER=groq
export GROQ_API_KEY={api_key}
# Execute screening
python scripts/03_screen_papers.py \
--project {project_path} \
--question "{research_question}" \
--max-workers 8 \
--batch-size 50
I2 validates AI evidence quotes against abstracts:
def validate_evidence_grounding(quotes, abstract):
"""Flag potential hallucinations"""
for quote in quotes:
if quote.lower() not in abstract.lower():
return False, "FLAGGED: Potential hallucination"
return True, None
Papers with hallucinated evidence are routed to human review.
| Keywords (EN) | Keywords (KR) | Action |
|---|---|---|
| screen papers, PRISMA screening | 논문 스크리닝, 선별 | Activate I2 |
| inclusion criteria, exclusion | 포함 기준, 제외 기준 | Activate I2 |
| AI screening, automated screening | AI 스크리닝 | Activate I2 |
I2 can call B2-evidence-quality-appraiser for deeper quality assessment:
Task(
subagent_type="diverga:b2",
model="sonnet",
prompt="""
Assess quality of included papers using:
- Risk of Bias (RoB) for RCTs
- Newcastle-Ottawa for observational
- GRADE for overall evidence quality
"""
)
requires: ["I1-paper-retrieval-agent"]
sequential_next: ["I3-rag-builder"]
parallel_compatible: ["B2-evidence-quality-appraiser"]
还没有评价。欢迎分享你的使用感受。
无需登录即可发表。欢迎分享使用感受、提问，或说明在哪些环境下无法运行。
还没有评论。来写第一条吧。
Turn a broad research idea into clear, testable questions. Compare familiar and less conventional approaches, structure the selected question around its population and methods, and assess variables, feasibility, and potential theoretical contributions.
中文简介正在准备中。当前显示英语简介。
Helps build a research framework by comparing theories, mapping relationships between variables, and deriving testable hypotheses. Explores less conventional options, examines weaknesses and competing explanations, and produces diagrams with anticipated reviewer objections.
中文简介正在准备中。当前显示英语简介。
Clarify the philosophical assumptions behind a research project and check whether they fit the proposed methods. Get a reasoned research stance, draft proposal wording, and guidance for explaining how the researcher's background and values shape the study.
中文简介正在准备中。当前显示英语简介。
Plan and draft academic papers through specialized agent roles covering literature, arguments, citations, and review. Get chapter guidance, revision roadmaps, Traditional Chinese and English abstracts, and documents formatted as LaTeX, DOCX, PDF, or Markdown.
中文简介正在准备中。当前显示英语简介。
Turn a detailed outline into a philosophy or interdisciplinary manuscript. Draft chapters with repeated quality reviews, check terminology and arguments across sections, and prepare the manuscript, evaluation reports, abstract, and metadata for submission.
中文简介正在准备中。当前显示英语简介。
Simulate academic peer review through five reviewer roles tailored to the paper's field. Receive separate critiques, an editorial decision, and a prioritized revision roadmap, or check whether a revised manuscript addresses earlier comments.
中文简介正在准备中。当前显示英语简介。