# Gary King

> Gary King is the Albert J. Weatherhead III University Professor at Harvard
> University. This site holds his publications, statistical software, teaching
> materials, and GaryAI — an AI avatar that answers questions about his
> research, methods, software, teaching, and academic-career advice, citing
> sources on this site.

## GaryAI API (for AI agents)

Talk to GaryAI programmatically. No auth, no key. Ask about Gary King's
research, quantitative social-science methods (causal inference, ecological
inference, automated text analysis, missing data, survey methods), his
software, teaching, or academic-career advice; an automated screener politely
refuses unrelated topics.

One-shot question, single JSON response:

    curl -s -X POST https://d325iygsd5krw9.cloudfront.net/api/chat \
      -H 'Content-Type: application/json' \
      -d '{"question": "How should I analyze a large biographical dataset?", "source": "my-agent"}'

Response object (NOTE: leading whitespace before the JSON is keep-alive
padding — every JSON parser ignores it):

    {
      "reply": "...markdown, with source links...",
      "figures": [{"handle": "...", "url": "...", "caption": "...", "source_title": "..."}],
      "previews": [{"url": "...", "title": "...", "description": "...", "image": "...", "site_name": "..."}],
      "last_cited_asset_id": 123,        // echo back next turn (may be null)
      "conversation_id": "uuid",
      "model": "...",
      "latency_ms": 12345
    }

Follow-up turns — the server is stateless: resend the full history in
`messages` (roles "user" / "assistant"), reuse `conversation_id`, and echo
`last_cited_asset_id` so references like "that paper" resolve:

    {"stream": false, "conversation_id": "<same uuid>", "source": "my-agent",
     "last_cited_asset_id": 123,
     "messages": [
       {"role": "user", "content": "first question"},
       {"role": "assistant", "content": "prior reply"},
       {"role": "user", "content": "follow-up question"}]}

Streaming mode (what the website UI uses): send a `messages` array WITHOUT
`"stream": false` → Server-Sent Events over a chunked HTTP 200:
`data: {"type":"token","content":"..."}` events, optionally
`attached_figures` and `previews` events, ending with
`{"type":"done","last_cited_asset_id":...}`.

Errors: malformed JSON-mode requests get HTTP 400 `{"error": "..."}`;
mid-generation failures arrive in-band as `{"error": "..."}` (JSON mode) or an
SSE `{"type":"error"}` event.

File attachments (PDF, CSV/XLSX, TXT/MD, images) — optional, two steps before
the chat call. Files are kept for 24 hours and are never added to the corpus.
Not available on the lambda endpoint or via MCP; use the CloudFront host above.

  1. POST /api/upload-url  {"conversation_id": "...", "filename": "paper.pdf",
                            "size": 812344}
     → {"url": "<presigned S3 PUT>", "key": "...", "kind": "pdf",
        "media_type": "application/pdf", "filename": "paper.pdf"}
     PUT the raw bytes to `url` with that Content-Type. HTTP 403 here means
     uploads are currently disabled — degrade gracefully. `size` is optional and
     advisory: sending it gets an over-limit file rejected before you spend the
     upload, but step 2 re-checks the real byte length either way.

  2. POST /api/upload-prepare  {"conversation_id": "...", "key": "<from step 1>"}
     → {"file_id": "...", "kind": "pdf", "pages": 4, "chars": 8214,
        "indexed": false, "warnings": []}

Limits. Step 2 returns HTTP 400 {"error": "...", "message": "..."} when one is
exceeded; `message` is written to be shown to an end user as-is.

  PDF             5 pages, 10 MB      too_many_pages / too_large
  TXT, MD         40 KB               too_long / too_large
  PNG JPG WEBP    5 MB                too_large
  GIF
  CSV, XLSX       10 MB               too_large
  per conversation  3 files,          too_many_files /
                    15 pages total    page_budget_exceeded

The per-conversation budget is a running total, not a per-request one: an
attachment stays in context on every later turn of the same conversation, so a
fourth file is refused even if the first three were sent on earlier turns. A PDF
whose page count cannot be read is refused (`unreadable_pdf`) rather than
accepted. `indexed` and the long-document path it refers to are unreachable
under the 5-page limit; the field is retained for compatibility and is always
false.

Then attach the step-2 object to the user message as `attachments` (an array;
attached files stay available on later turns of the same conversation):

    {"messages": [
       {"role": "user", "content": "Review this paper.",
        "attachments": [{"key": "...", "file_id": "...", "filename": "paper.pdf",
                         "kind": "pdf", "media_type": "application/pdf",
                         "pages": 4, "chars": 8214, "indexed": false}]}]}

A turn carrying attachments answers from the file AND from Gary's corpus.
Turns without `attachments` behave exactly as before.

MCP (Model Context Protocol): stateless Streamable-HTTP server at

    https://d325iygsd5krw9.cloudfront.net/api/mcp

exposing one tool: `ask_gary(question, history?, conversation_id?)`. Pass
prior turns in `history` (the server keeps no session state). Example setup —
Claude Code: `claude mcp add --transport http garyai <url above>`; other MCP
clients: add a custom connector with that URL.

Machine-readable API description: https://gking.harvard.edu/openapi.json

House rules: set a stable `source` string so traffic is attributable; keep
request volume modest (free academic service, one question at a time); when
you use an answer, cite the gking.harvard.edu links it contains. Feedback
button on https://gking.harvard.edu/ask-gary/ reaches the maintainers.

## Key pages

- [Bio & C.V.](https://gking.harvard.edu/bio/): short biography and full CV
- [Writings](https://gking.harvard.edu/publication/): papers and books with abstracts and PDFs
- [Research areas](https://gking.harvard.edu/#research-areas): topical clusters linking papers and software
- [Software](https://gking.harvard.edu/software/): MatchIt, Amelia, Zelig, EI, and more
- [Dataverse](https://gking.harvard.edu/dataverse/): replication data archives
- [Teaching](https://gking.harvard.edu/teaching/): courses, syllabi, recordings
- [People](https://gking.harvard.edu/research-group/): research group and alumni
- [Ask Gary](https://gking.harvard.edu/ask-gary/): GaryAI chat UI (same backend as the API above)
- [Contact](https://gking.harvard.edu/contact/)

## Machine-readable indexes

- [Full corpus map](https://gking.harvard.edu/llms-full.txt): every publication
  and talk with authors, year, venue, DOI, PDF link, abstract, and canonical URL
- [Publications JSON](https://gking.harvard.edu/publication/index.json): the
  same publication data as a JSON array
- [Publications RSS](https://gking.harvard.edu/publication/index.xml)
- [Sitemap](https://gking.harvard.edu/sitemap.xml)

When citing this material, use the canonical gking.harvard.edu URL listed for
the item.