InternScience/sciverse-academic-retrieval

Sciverse Academic Retrieval

Connects to the Sciverse SCP Server via the SCP Hub MCP gateway to perform citation-grade scientific literature retrieval over a corpus that includes peer-reviewed papers (Nature, Cell, …), preprints (arXiv, bioRxiv, …) and other academic sources.

The server exposes 5 tools designed for RAG by autonomous research agents: structured metadata search, natural-language semantic chunk retrieval, byte-range source-text reading, and figure/table image fetching — all returning stable doc_id / chunk_id for reproducible citation.

Usage

import asyncio
import json
import base64
from mcp.client.streamable_http import streamablehttp_client
from mcp import ClientSession


class SciverseClient:
    """Sciverse SCP Server client (5 academic-retrieval tools).

    All requests transparently proxied by the SCP Hub to the Sciverse backend.
    Authentication uses the SCP-HUB-API-KEY header (your SCP Platform key).
    """

    def __init__(self, server_url: str, api_key: str):
        self.server_url = server_url
        self.api_key = api_key
        self.session = None

    async def connect(self):
        try:
            self.transport = streamablehttp_client(
                url=self.server_url,
                headers={"SCP-HUB-API-KEY": self.api_key},
            )
            self.read, self.write, self.get_session_id = await self.transport.__aenter__()
            self.session_ctx = ClientSession(self.read, self.write)
            self.session = await self.session_ctx.__aenter__()
            await self.session.initialize()
            return True
        except Exception as e:
            print(f"[sciverse] connect failed: {e}")
            return False

    async def disconnect(self):
        if self.session:
            await self.session_ctx.__aexit__(None, None, None)
        if hasattr(self, "transport"):
            await self.transport.__aexit__(None, None, None)

    def parse_text_result(self, result):
        """Extract concatenated text from a tool result's content blocks.

        Works for: search_papers, semantic_search, read_content, list_catalog.
        Returns: str (the tool's JSON payload as text).
        """
        if isinstance(result, dict):
            content_list = result.get("content") or []
        else:
            content_list = getattr(result, "content", []) or []
        texts = []
        for item in content_list:
            if isinstance(item, dict):
                if item.get("type") == "text":
                    texts.append(item.get("text") or "")
            else:
                if getattr(item, "type", None) == "text":
                    texts.append(getattr(item, "text", "") or "")
        return "".join(texts)

    def parse_image_result(self, result):
        """Extract a figure/table image (used by get_resource).

        Returns: dict with keys 'mime_type' (e.g. 'image/png') and 'bytes'
                 (decoded binary). Returns None if the result is not an image.
        """
        if isinstance(result, dict):
            content_list = result.get("content") or []
        else:
            content_list = getattr(result, "content", []) or []
        for item in content_list:
            data = item.get("data") if isinstance(item, dict) else getattr(item, "data", None)
            mime = item.get("mimeType") if isinstance(item, dict) else getattr(item, "mimeType", None)
            type_ = item.get("type") if isinstance(item, dict) else getattr(item, "type", None)
            if type_ == "image" and data:
                return {"mime_type": mime, "bytes": base64.b64decode(data)}
        return None

Initialize and use

SERVER_URL = "https://scp.intern-ai.org.cn/api/v1/mcp/43/Sciverse"
API_KEY = "<YOUR_SCP_HUB_API_KEY>"


async def main():
    client = SciverseClient(SERVER_URL, API_KEY)
    if not await client.connect():
        print("connect failed")
        return
    try:
        # 1. Structured search: recent transformer papers
        result = await client.session.call_tool(
            "search_papers",
            arguments={
                "query": "transformer attention",   # BM25 over title/abstract/journal
                "year_from": 2023,
                "page_size": 5,
            },
        )
        papers = json.loads(client.parse_text_result(result))
        print(f"search_papers hits: {len(papers.get('hits', []))}")

        # 2. Semantic search: RAG-style chunk retrieval
        result = await client.session.call_tool(
            "semantic_search",
            arguments={"query": "How does transformer attention work?", "top_k": 3},
        )
        chunks = json.loads(client.parse_text_result(result))
        for hit in chunks.get("hits", []):
            print(f"  - {hit['title']} (score={hit['score']:.3f}, doc_id={hit['doc_id']})")

        # 3. Read content: expand context around a known offset
        if chunks.get("hits"):
            first = chunks["hits"][0]
            result = await client.session.call_tool(
                "read_content",
                arguments={"doc_id": first["doc_id"], "offset": first["offset"], "limit": 4096},
            )
            text_window = json.loads(client.parse_text_result(result))
            print(f"read_content next_offset={text_window.get('next_offset')} more={text_window.get('more')}")

        # 4. List catalog: discover available filter fields and operators
        result = await client.session.call_tool(
            "list_catalog", arguments={"include_sample_values": False},
        )
        catalog = json.loads(client.parse_text_result(result))
        print(f"available filter fields: {len(catalog.get('fields', []))}")

        # 5. Get resource: fetch a figure referenced inside read_content's Markdown
        # (Only call after read_content returned a Markdown snippet with ![alt](file_name).)
        # result = await client.session.call_tool(
        #     "get_resource", arguments={"file_name": "figures/fig-3.png"},
        # )
        # image = client.parse_image_result(result)
        # if image:
        #     from pathlib import Path
        #     Path("fig-3.png").write_bytes(image["bytes"])
    finally:
        await client.disconnect()


asyncio.run(main())

Tool: search_papers

Structured metadata search by author, journal, year, subject, etc. Use when the user knows specific filter values ("Hinton's papers from 2020-2023", "Nature papers on CRISPR"). Do not use for free-text Q&A — that's semantic_search.

• Args:
  • query (str, optional) — BM25 keyword over title/abstract/journal
  • title_contains (str, optional) — Substring match on title
  • abstract_contains (str, optional) — Substring match on abstract
  • authors (list[str], optional) — Any of these authors matches
  • year_from / year_to (int, optional) — Publication year range (inclusive)
  • journals (list[str], optional) — Journal names (any match)
  • subjects (list[str], optional) — Subject classification (e.g. "biology")
  • sort_by_year (str, default "desc") — desc / asc / none
  • page (int, default 1), page_size (int, default 10, max 50)
  • filters_advanced (list, optional) — Escape hatch with full operator set (FILTER_OP_EQ, IN, CONTAINS, GTE, LTE, …) for fields not surfaced above
• Returns: JSON {hits: [...], total: int} where each hit has doc_id, title, author, abstract, publication_venue_name, publication_published_year.

Tool: semantic_search

Natural-language semantic search returning relevant paper chunks for RAG-style answering. Use for free-text questions ("How does attention work?"). Typical chain: semantic_search → pick chunk → read_content.

• Args:
  • query (str, required) — Natural-language query, 1-200 words optimal
  • top_k (int, default 10, max 30)
  • source_types (list[str], optional) — Filter by web / pdf
  • mode (str, default "balanced") — fast (~200ms keyword only) / balanced (~600ms hybrid) / quality (~2-4s LLM-rewrite + hybrid)
• Returns: JSON {hits: [...]} where each hit has chunk_id, doc_id, chunk (the matched text), score, title, offset (byte offset into source doc — pass to read_content for expansion).

Tool: read_content

Read a UTF-8 byte range of a paper's source text. Typically called with a doc_id/offset returned by semantic_search to expand context (read more bytes before or after a chunk for fuller answers).

• Args:
  • doc_id (str, required) — Paper ID from search_papers / semantic_search
  • offset (int, default 0) — Byte offset to start reading
  • limit (int, default 4096, max 16384) — Bytes to read
• Returns: JSON {text: str, bytes_returned: int, next_offset: int, more: bool}. Markdown text may contain figure references like ![alt](file_name) — pass file_name to get_resource to fetch the image.

Tool: get_resource

Fetch the binary bytes of a paper figure / table image referenced inside read_content's Markdown. Use when the user asks to see / describe a figure and read_content output contains an image reference.

• Args:
  • file_name (str, required) — Relative path from the Markdown ![alt](file_name). Must not contain .. or start with /.
• Returns: Image content block — data (base64) + mimeType (image/*). Multimodal agents (Claude, GPT-4V, Gemini, …) can read it directly.

Tool: list_catalog

Returns the schema catalog for search_papers: every field name, type, whether it's filterable / sortable / default-returned, human description, and applicable filter operators. Use when constructing precise search_papers filters or facing an ambiguous field need.

• Args:
  • include_sample_values (bool, default false) — If true, also fetch top-20 values for enum-like fields (24h cached, ~100s of ms first call).
• Returns: JSON {fields: [...]} where each field has name, type (string/integer/list[string]/…), filterable, sortable, default_return, description, applicable_operators, and optionally sample_values.

Use Cases

• Drug discovery / pharmacology: literature scoping for a target before triggering wet-lab skills; RAG context for ADMET / MoA reasoning.
• Protein science: gather structure/function papers around a UniProt ID before predicting mutations or binding sites.
• Genomics & rare disease: pull recent papers on a variant / phenotype for evidence-grade reasoning, then cite by doc_id.
• Chemistry / materials: find prior art around a SMILES or reaction before computing properties.
• Cross-domain literature review: agentic survey writing — chain semantic_search → read_content to assemble citation-grounded summaries with stable doc_id references for verifiability.