Session library

sessions.dataimago.ai/<conference-year>/<session>: the shared, citable materials of conference sessions, for people (browse, search, ask with citations) and for AI agents (MCP, REST, llms.txt). Read-only, no login.

A company-operated Level-2 surface of the sesh vertical (not a user-owned Level-3 app). Decision of record: sesh-ai/wiki/decisions/sessions-library.md.

Where content comes from

Two pinned inputs, nothing else:

• Documents: from a public repository the session's participants chose to share. Its bundle manifest lists the citable files, which are read from the git objects of one commit.
• Session frame (title, participants, guiding questions, run of show, i.e. the published program): from the session's committed SessionSpec in sesh-ai, pinned by hash.

sesh-ai's pnpm corpus:bundle writes both into content/<conference-year>/<session>/bundle.json (corpus bundle v1) with their provenance (commitSha, manifestSha256, specSha256). This app never reads lain's corpus copy, sesh-ai's database, store/, recordings/ or exports/.

Collection	Source
aime-2026/ai-native-profession	dbetebenner/AIME_2026_Panel_Presentation · corpus-manifest.yml

Add or refresh a session

Commit the manifest (in the source repository) and the SessionSpec (in sesh-ai) first: the bundler reads committed bytes only and refuses uncommitted changes, symlinks, staff-owned documents, and an --out under store/, recordings/, exports/ or fixtures/.

# in sesh-ai: build the bundle from the committed manifest
pnpm corpus:bundle --manifest <repo>/corpus-manifest.yml \
  --spec specs/<session>.sessionspec.yaml \
  --out ../sessions/content/<conference-year>/<session>/bundle.json

# here: embed new chunks (needs OPENAI_API_KEY), regenerate the registry, test
pnpm index
pnpm registry
pnpm test && pnpm build

Commit bundle.json and embeddings.json. Builds and deploys never call a model.

Surfaces

• Pages: / · /<conference> · /<conference>/<session> (panel, questions, documents) · /documents/<id> (one anchor per passage) · /search · /ask
• REST: /api/v1/collections/... (see /api/openapi.json, /api/discover)
• MCP: /api/mcp, streamable HTTP, stateless, read-only. Tools: list_collections, list_documents, get_document, search_corpus, get_passage, get_provenance. No ask tool. pnpm mcp:self-test <url> checks a running server. Claude Code: claude mcp add --transport http sessions https://sessions.dataimago.ai/api/mcp

Ask

Claude Sonnet 5.5 answers only from retrieved passages, labels each statement retrieved or inferred, cites passage ids, and abstains (naming who would know) when the materials do not answer. It is not lain. Citations are checked server-side. Limits: 4 questions per minute and 30 per day per IP; a global daily cap (SESSIONS_ASK_DAILY_CAP_USD, default $3) reserved before each call. Only outcome counts are recorded, never questions. In production Ask refuses without Upstash Redis, because a per-instance cap is not a cap.

Environment

See .env.example for the names.

Deploy

• Vercel project sessions (team dataimago-projects), connected to this repository. Merging to main deploys production; every pull request gets a preview.
• Domain: sessions.dataimago.ai, a CNAME at the registrar to the value Vercel shows under Settings → Domains.
• Upstash Redis (Vercel Marketplace, pay as you go) supplies KV_REST_API_URL and KV_REST_API_TOKEN for the rate limits and the Ask cap.
• Model keys: ANTHROPIC_API_KEY and OPENAI_API_KEY, set by the maintainer as sensitive variables for Production and Preview. SESSIONS_ORIGIN (https://sessions.dataimago.ai, Production only) and SESSIONS_ASK_DAILY_CAP_USD (3) are plain settings.
• Checking a preview: previews sit behind Vercel Authentication, so plain curl gets a redirect. Use vercel curl <path> --deployment <preview-url>, which passes the logged-in session.
• Reviews: CI (lint, typecheck, test, build, registry drift) and a Grok 4.6 review on every pull request, with rules in .cursor/BUGBOT.md. The maintainer merges.