Skip to content

Latest commit

 

History

History
133 lines (101 loc) · 5.32 KB

File metadata and controls

133 lines (101 loc) · 5.32 KB

Base2026 API & AI Access

Base2026 exposes the same reviewed public evidence to people and software. The public layer is read-only: source metadata, short evidence passages, public cards, topics, attribution, and canonical links.

Raw captions, raw ASR, media, private review packets, credentials, logs, and private pipeline state are not part of the public API.

Public entry points

  • Human search workspace: /workspace/
  • Public dataset landing page: /dataset
  • Agent-readable context: /llms.txt
  • Public data dictionary: /data-dictionary.json
  • Public API index: /api-index.json
  • Static sitemap index: /sitemap.xml
  • Current D1 projection sitemap: /sitemap-dynamic.xml
  • Public source index: /sources/

Static public data

  • /static/manifest.json — dated release counts and public export metadata.
  • /static/documents.jsonl — public source/search documents.
  • /static/passages.jsonl — public evidence passages linked to sources.
  • /static/insight_cards.jsonl — reviewed source-backed insight cards.
  • /static/topic_signal_briefs.jsonl — summaries for strong public topics.

Static files are best for reproducible offline analysis. Their manifest date must be retained when citing counts.

The public dataset landing page links these distributions to the live D1 layer, source policy, machine-readable catalog, and a copy-ready API query.

Live search endpoint

The public UI and compatible integrations use a read-only, Meilisearch-compatible Cloudflare Worker endpoint backed by D1 FTS5:

POST /api/search/multi-search

No browser key is required. The endpoint provides public search and filters; it has no write, moderation, raw-transcript, media, credential, or private-data route.

Example:

{
  "queries": [
    {
      "indexUid": "base2026_public_tiktok",
      "q": "AI search",
      "limit": 5
    }
  ]
}

Current read-only Worker routes

The same public boundary also covers these bounded, read-only routes:

  • GET /api/health for Worker and public-search liveness;
  • GET /api/stats for current public corpus and privacy-boundary totals;
  • GET /api/evidence-brief?q=... for deterministic Evidence Brief V1;
  • GET /api/evidence-brief/v2?q=... for bounded attributable findings with corpus and ranking receipts;
  • GET /api/blog and GET /api/blog/{slug} for approved editorial metadata;
  • GET /api/guides and GET /api/guides/{slug} for maintained task guides whose public-source dependencies are checked before serving.
  • GET /api/claim-receipts/v1?canary=base2026.internal-linking.canary.v1&topic=internal-linking for the bounded ten-receipt internal-linking canary. The route is held with 503 CLAIM_RECEIPT_CANARY_NOT_READY until public D1 contains exactly ten validated active receipts; it has no public write or correction method.

Editorial and guide records are separate from source-corpus counts. No public route writes, approves or moderates content.

The claim-receipt schema, deterministic static sidecars and owner-only admission boundary are documented in the claim-receipt ledger contract.

Indexable public projection

Eligible automatic D1 projections receive stable public source pages at:

/sources/tiktok-video-{numeric_video_id}

These pages show only sanitized public excerpt cards and attribution. The original creator video remains the canonical source for the full content.

Good uses

  • find creators and sources discussing a topic;
  • compare repeated tactics across public sources;
  • inspect source-backed SEO, GEO, AEO, and AI-search claims;
  • build a research notebook from public JSONL;
  • link an answer to a stable Base2026 page and the original source.

Not supported

  • raw transcript harvesting or video re-hosting;
  • creator impersonation;
  • private lead, inbox, review, or administrative access;
  • public writes, corrections, or moderation through the search endpoint;
  • replacing the original creator channel.

Read-only MCP contract

POST /api/mcp is a stateless JSON-RPC surface over public D1. It supports the current MCP discovery and legacy 2025-11-25 initialization compatibility, with bounded calls to:

  • search_sources for public source/evidence lookup;
  • get_source for one canonical source record;
  • get_creator for creator metadata and linked public sources;
  • get_topic for a topic and its public source summary;
  • get_topic_signal for a deterministic public topic signal;
  • get_public_manifest for dated public release dimensions.

The route has no sessions, SSE, writes, moderation, private bindings or credentials. Responses preserve attribution, original source links, canonical Base2026 URLs and public/private policy flags. Request bodies, arguments and returned evidence are bounded; raw captions, raw ASR, full private transcripts, media, inbox data and pipeline control state are never returned.

The Worker configuration includes a MCP_RATE_LIMIT binding at 60 requests per minute per edge identity. Exhausted clients receive 429 with a retry hint; when the binding is missing or unavailable, the route fails closed with 503. Production releases verify the binding through Cloudflare version readback and exercise the endpoint after deployment. Use the MCP guide and integration guide for request examples and client setup.