ATLAS

AI Taxonomy, Literature Analysis & Survey — a research atlas for discovering AI papers, exploring their taxonomy and reading structured analyses. The website brand is AI-tlas; the repository name is ai-taxonomy-literature-analysis-survey.

Explore

• /: landing page with paper search, live collection counts and topic entry points.
• /overview: venue coverage, conference dates and topic counts (the former homepage).
• /papers and /papers/[id]: search/filter papers and read their analyses.
• /landscape: topic taxonomy, work types and research landscape.
• /contents: surveys and analysis groups, including tools of agents and Lean theorem proving.
• /work-types/[slug], /benchmarks/[id], /conferences/[slug]: detailed research views.

The current collection focuses on available 2026 NLP/ML venue lists. Sources include ICLR/ICML through papercopilot/paperlists and ACL-family venues through ACL Anthology. Availability depends on upstream sources; this is not a complete index of all AI research.

Stack and storage

Next.js Pages Router, React, TypeScript and PostgreSQL (pg). Runtime requires DATABASE_URL; lib/db.ts uses search_path=live.

• live: 14 imported tables for conferences, papers, taxonomies, work types, labels, benchmarks, results and reviews. Loading builds a temporary next schema, verifies row counts and atomically replaces live. Treat these tables as replaceable imported data.
• analysis: persistent editorial content. page_block stores ordered typed JSONB blocks keyed by paper/<conf>/<ext_id> or content/<node>. Tool categories, paper tools and category mappings also live here. Legacy article/section/block/link tables remain for migration; current pages render page blocks. Rebuilding live does not replace analysis.
• Repository data: data/labels/o*.jsonl contains 2,678 paper labels; data/labels/type_notes.json contains per-type syntheses. These are still tracked import sources, not database-only content. scripts/worktypes.json, scripts/taxonomy.py, lib/content-tree.ts and lib/content-defaults.ts define vocabulary, classification rules, content navigation and initial blocks.
• Files and media: raw downloads, staging SQLite databases and exports are ignored by Git. Deployment caches raw sources and JSONL exports under /opt/llm-survey/cache/. Figure blocks store a source URL, alt text and caption, not image bytes. Charts and Mermaid are rendered by components. No image storage/upload service is implemented.

Schema definitions: scripts/load-pg.js (live) and lib/analysis-schema.ts (analysis). Use production database catalogs to verify actual deployed structure.

Data quality

Topic assignment uses keyword scoring: title ×3, keywords ×2, abstract ×1. Agent-paper labels use titles and abstracts; other papers use rule-based labels. Scores are those stated in abstracts. Labels, summaries and inferred classifications should be checked against the original papers.

CI/CD and daily refresh

Gitea Actions configuration: .gitea/workflows/deploy.yml; deployment: scripts/deploy.sh.

• Pushes to main trigger deployment, except documentation-only changes (**/*.md, docs/**).
• A job-level branch guard restricts deployment to refs/heads/main, including manual dispatches. Feature branches and PRs do not deploy.
• A scheduled run at 03:17 Asia/Shanghai daily (19:17 UTC) requests fresh upstream data, rebuilds the data export, loads PostgreSQL when changed, builds the site and restarts/checks the service. The schedule runs from the default branch; keep that branch set to main.
• Manual dispatch supports refresh_data: false reuses cached data where possible; true downloads and rebuilds data.
• Feature-branch pushes run type checking and a production build on ces-control via .gitea/workflows/validate.yml; validation never deploys or imports data.
• The deployment runner is ces-control. Building and deployment occur there, not on a developer's local machine. Scheduler/runner availability determines whether a scheduled run actually succeeds; inspect Gitea Actions logs for results.
• Production service remains llm-survey, under /opt/llm-survey, with database llm_survey, environment /etc/llm-survey/db.env, and default port 18231. These infrastructure identifiers are intentionally stable across the branding change.

The deployment cache reuses unchanged exports, dependencies and Next.js compilation cache. A changed export replaces live; persistent page blocks and tool annotations in analysis are retained. Back up both schemas before changing data ownership or migrations.

Remote development and validation

Install dependencies and run npm run typecheck / npm run build only in the approved remote execution environment. For an isolated remote preview, set DATABASE_URL and run npm run dev; do not point a feature-branch deployment at the production service.

Data pipeline commands (remote runner only):

npm run data:fetch
npm run data:build
npm run data:load

data:build stages SQLite under build/ and exports JSONL; data:load imports into PostgreSQL and replaces live. These commands change data and are not read-only validation.

Public journal and editorial access

The public site now presents only the two initial studies at /contents/lean-theorem-proving and /contents/tools-of-agents. / retains live collection counts, and /contents lists those studies. Other paper, benchmark, conference and landscape routes require an editor session. Their database rows are preserved; this change does not delete or reimport content.

• /editor/login: a six-digit authenticator TOTP code (no password).
• /editor: editorial desk and private archive entry.
• /editor/contents/<study>: edit text blocks directly (JSON mode for structured content), preview a draft, then explicitly publish. Preview resolves dynamic blocks without saving. Publishing replaces the study's blocks immediately; this first editor has no durable drafts or per-editor accounts. Coordinate edits to avoid overwrites.
• All content and taxonomy APIs, including GET/list endpoints, require the signed editor cookie. Mutations require the configured Origin. API clients must sign in through the same authenticator flow; there is no anonymous write fallback.

Configure these secrets in /etc/llm-survey/db.env through your server secret management process, never in Git or CI logs:

• EDITOR_ORIGIN: the exact browser origin, including port (no trailing slash). Use HTTPS for internet-facing editorial access; HTTPS enables Secure cookies.
• EDITOR_TOTP_SECRET: a random base32 authenticator secret (at least 32 base32 characters), enrolled privately in the editor's authenticator. TOTP uses SHA-1, 30-second periods and six digits; server time must be synchronized.
• EDITOR_SESSION_SECRET: at least 32 random characters for HMAC-SHA256 sessions. Rotating this value revokes all sessions.

Sessions expire after eight hours, are HttpOnly and SameSite=Strict, and can be ended from the desk. Login attempts are globally limited to ten per five minutes in persistent analysis.editor_auth; accepted TOTP time steps cannot be reused. Missing configuration leaves editing closed while public studies remain readable. This is a single shared editorial identity; secret rotation is the recovery procedure. The existing HTTP public URL can continue serving readers; configure TLS and the editor secrets before using editorial login over the internet.

The authenticator QR is provisioned once and delivered privately in chat, never committed. TOTP-only login is a single possession factor, as requested.

The deployment job can provision EDITOR_TOTP_SECRET and EDITOR_SESSION_SECRET from repository Actions secrets after explicit operator authorization. It writes them atomically to the existing environment file with mode 0600. Missing secrets preserve the server configuration. Update EDITOR_ORIGIN in the deployment job when moving editorial access to a TLS domain.