Build a Literature Review Agent with Sciverse

Start from one research question, retrieve evidence, summarize papers, and generate a cited literature review

Scenario

Researchers or AI agents need to automatically retrieve relevant papers, extract evidence passages, and generate a cited literature review for a research question.

Estimated calls

~15-30 API calls per review task

Tools

agentic-searchcontent

Pipeline

agentic-search→ doc_id + chunk + offset→ content→ evidence markdown

Input example

User asks in Claude / Cursor:
"Please review the applications of Transformers in protein structure prediction from 2020 to 2024, and list key papers and core contributions."

Output example

## Literature Review: Transformer Applications in Protein Structure Prediction (2020-2024)

### 1. The AlphaFold2 breakthrough
Jumper et al. (2021) proposed AlphaFold2, using the Evoformer module...
[Source: Nature, doc_id: af2_xxx, offset: 12480]

### 2. End-to-end prediction with ESMFold
Lin et al. (2023) proposed ESMFold...
[Source: Science, doc_id: esm_yyy, offset: 8320]

---
Retrieved 12 core papers and extracted 28 evidence passages.

Agent Prompt example

You are a scientific literature review agent. When the user provides a research question:
1. Call agentic-search(query=user question, top_k=20) to retrieve relevant passages
2. For each high-scoring passage, call content(doc_id=hit.doc_id, offset=hit.offset, limit=2000) to read context
3. Organize a structured review, and cite every claim with [doc_id, offset]
4. Do not fabricate references; every statement must be grounded in Sciverse results

Implementation steps

Step 1: Set up the environment

Install dependencies and configure API tokens

# 安装所需 Python 包
!pip install httpx anthropic
# 设置环境变量（替换为你的真实 Token）
import os
os.environ["SCIVERSE_API_TOKEN"] = "sv-your-token-here"  # 替换为你的真实值
import os
os.environ["ANTHROPIC_API_KEY"] = "sk-ant-..."  # 替换为你的真实值

Step 2: Retrieve relevant passages semantically

Use agentic-search to retrieve paper passages most relevant to the research question

import os
import asyncio
import httpx

BASE = "https://api.sciverse.space"
TOKEN = os.environ["SCIVERSE_API_TOKEN"]
HEADERS = {"Authorization": f"Bearer {TOKEN}"}

async def search_literature(query: str, top_k: int = 20):
    async with httpx.AsyncClient(timeout=30) as client:
        resp = await client.post(
            f"{BASE}/agentic-search",
            headers=HEADERS,
            json={"query": query, "top_k": top_k}
        )
        resp.raise_for_status()
        return (resp.json().get("hits") or [])

async def main():
    hits = await search_literature(
        "Transformer applications in protein structure prediction 2020-2024"
    )
    print(f"Found {len(hits)} relevant chunks")
    for h in hits[:3]:
        print(f"  [{h['score']:.2f}] {h['title'][:60]}...")
    return hits

hits = await main()

Step 3: Read source context

Call the content endpoint for high-scoring passages to read broader context

async def read_context(doc_id: str, offset: int = 0, limit: int = 2000):
    """读取指定文档的原文片段。返回 {text, next_offset, more}"""
    async with httpx.AsyncClient(timeout=30) as client:
        resp = await client.get(
            f"{BASE}/content",
            headers=HEADERS,
            params={"doc_id": doc_id, "offset": offset, "limit": limit}
        )
        resp.raise_for_status()
        return resp.json()

async def gather_evidence(hits, top_n=5):
    sorted_hits = sorted(hits, key=lambda x: x["score"], reverse=True)[:top_n]
    evidences = []
    for hit in sorted_hits:
        ctx = await read_context(hit["doc_id"], hit.get("offset", 0))
        evidences.append({
            "title": hit["title"],
            "doc_id": hit["doc_id"],
            "offset": hit.get("offset", 0),
            "chunk": hit.get("chunk", ""),
            "context": ctx["text"],  # 注意：响应字段是 text
            "score": hit["score"]
        })
    return evidences

evidences = await gather_evidence(hits)

Step 4: Generate a cited review (optional enhancement)

Use an LLM to synthesize the extracted evidence into a structured review

from anthropic import Anthropic

client = Anthropic()  # 自动读取 ANTHROPIC_API_KEY

evidence_text = "\
\
".join([
    f"[{e['doc_id']}, offset={e['offset']}] {e['title']}\
{e['context']}"
    for e in evidences
])

msg = client.messages.create(
    model="claude-sonnet-4-20250514",
    max_tokens=4096,
    messages=[{
        "role": "user",
        "content": f"""基于以下文献证据，生成一份关于 Transformer 在蛋白质结构预测中应用的综述。
每个论点必须标注来源 [doc_id, offset]。
不要编造任何未在证据中出现的信息。

{evidence_text}"""
    }]
)
print(msg.content[0].text)

Notes

Do not cite papers that were not returned by Sciverse.
Keep doc_id and offset in every evidence record so users can trace each claim.
Use content to read more context before summarizing when a retrieved chunk is too short.
Use top_k carefully; larger values improve recall but increase latency and downstream LLM cost.
For formal writing, keep evidence extraction and final writing as two separate steps.

FAQ

适合什么综述场景?

适合从开放研究问题出发生成初版文献综述，并保留引用证据。

如何保证综述可追溯?

每个结论都应绑定 doc_id、quote 和 offset，不能只输出模型总结。

需要人工复核吗?

需要，Agent 适合生成初稿和证据包，最终观点仍应由研究者确认。

和普通搜索结果有什么区别?

本案例强调证据链、上下文和引用位置，而不是只返回论文列表。

Next steps

Need an API key?

Create one in Console > Tokens.The same API key works for enabled Sciverse, DianShi, and Skills capabilities, with starter quota available according to account permissions.

Open console