Scientific database connectors

OpenAI4S exposes a thin, schema-normalized layer over common public scientific APIs. It keeps the tool surface flat while letting Python cells combine real records without scraping provider-specific pages.

Supported databases

The public model-facing surface remains two tools:

• science_list_dbs(domain?) lists sources, query hints, filters, and the normalized result contract.
• science_search(database, query, limit?, cursor?, filters?, timeout?) searches one source and returns typed records.

Every result uses the same envelope:

{
  "database": "uniprot",
  "source": "UniProtKB",
  "query": "insulin",
  "count": 1,
  "results": [
    {
      "id": "P01308",
      "title": "Insulin",
      "url": "https://www.uniprot.org/uniprotkb/P01308",
      "type": "protein",
      "attributes": {"gene_names": ["INS"], "taxon_id": 9606}
    }
  ],
  "next_cursor": null,
  "request_url": "https://rest.uniprot.org/..."
}

Code-cell composition

The same two operations are available through the injected singleton. This is the intended path when several pages must be joined, filtered, or analyzed in one persistent cell:

sources = host.science.list_databases("chemistry")
aspirin = host.science.search("pubchem", "aspirin", limit=5)
papers = host.science.search(
    "openalex",
    "aspirin pharmacogenomics",
    limit=25,
    filters={"year_from": 2022, "work_type": "article"},
)
rows = [
    {"cid": aspirin["results"][0]["id"], **paper}
    for paper in papers["results"]
]

Source-specific filters are intentionally bounded:

• organism_id for UniProt;
• species for an exact Ensembl gene-symbol lookup (default homo_sapiens);
• year_from, year_to, and work_type for OpenAlex.

arXiv and OpenAlex return cursors. Other first-batch connectors are bounded single-page searches. PubChem uses its exact name/synonym endpoint rather than claiming fuzzy text-search semantics.

Safety and failure behavior

Connectors construct URLs from fixed HTTPS endpoints; callers cannot supply a host or arbitrary URL. Requests use the existing Web fetch path, so the global network switch, per-redirect SSRF checks, egress allowlist, response-size cap, timeouts, permission audit, and prompt-injection annotation remain active. Network/API/schema failures return the normal single-key { "error": "..." } soft-fail shape at the tool boundary. No connector adds a runtime dependency to the stdlib-only core, and all default tests use captured synthetic API payloads without network access.

Doubao Search Custom

Doubao Search Custom is the primary managed web-search option. OpenAI4S calls Volcengine's fixed search API through a stdlib-only client and normalizes web hits to the exact {title,url,snippet} envelope. It does not install the Ark CLI Skills catalog or a local third-party MCP package.

The client resolves the same Agent Plan Key used by Ark and DataPro from SecretBroker immediately before each outbound request. The credential is sent only as the upstream Authorization value and is never returned by the config route, search response, DOM, or diagnostic text. Saving it from Customize → Network therefore authorizes both managed products once; an active Ark API key is reused only when the selected provider is Ark.

The dedicated POST /doubao-search/search product check has no fallback. It does not call Tavily, DuckDuckGo, Bing, Mojeek, or an identifier resolver after an upstream failure or empty result. The server and UI report “豆包搜索可用” only when a real response is identified as source: "doubao" and contains at least one normalized result with a non-empty URL. Saving a key, opening the endpoint, or receiving an empty success response is not an authentication or readiness verdict. Tavily remains visible in Customize → Network as a backup option for the separate generic search path.

Volcengine DataPro professional datasets

DataPro is intentionally separate from the normalized public-database tools above. It is one fixed, managed MCP Streamable HTTP connector named volcengine-datapro, and exposes only dataPro_search(query:string) to its bundled Skill. The endpoint is not user-selectable. The Agent Plan Key is stored through SecretBroker and resolved only as each outbound POST is assembled; an active Ark model key is reused only when the active provider is Ark.

The Customize → Connectors card performs the actual search call and persists its result as a JSON Artifact. Every successful response is indexed without a field allowlist: all keys, scalar values, nested objects, nested arrays, MCP content blocks, text, future result-envelope fields, and duplicate logical occurrences in the redacted result returned by that call are covered by a leaf-count and digest completeness receipt. This is a guarantee about all content returned by that query, not a claim that the single-argument MCP tool can enumerate DataPro's entire remote corpus.

Neither MCP initialization nor tool discovery is an authentication verdict. The UI reports “专业数据集可用” only after the real tool result contains an integer structuredContent.code equal to 0 and the local index transaction returns a complete receipt with matching source/index leaf counts; code 4011 reports “Key 无效、额度不足，或者专业数据集 Harness 未开启。”. The dedicated Web route, managed connector call, and bundled Skill's host.mcp.call path share this ingestion boundary. Indexed entries are discoverable from the global command palette and link back to their saved Artifact when one exists. The default permission seed allows only the exact volcengine-datapro/dataPro_search target. Supplying or activating the shared brokered Agent Plan credential is the user's authorization; the bundled connector and Skill are enabled by default, and the UI enable action is idempotent. All other MCP calls keep the existing ask policy, and any operator deny remains an absolute veto. Redirects are refused so authenticated headers cannot cross origins, and the fixed endpoint still passes the global network switch, exact-host egress allowlist, SSRF guard, timeout, and response-size ceiling.