#Data reference

Field-level documentation for the machine-readable formats. If you are writing code against an export, this is the page you want.

• JSON
• JSONL
• CSV
• Library manifest
• Value glossary

#JSON

schemaVersion is 1. A breaking change to any field below bumps it; adding an optional field does not.

Fields marked optional are omitted entirely, not set to null, when the setting that controls them is off. Read them with a default. The distinction is deliberate: undefined means "not exported", null means "AI Studio had no value".

#Top level

#config

null when AI Studio reported no configuration; individual fields are null when that one value was absent.

#stats

Recomputed after settings are applied, so it describes what is in the file, not what the conversation originally held. Exclude reasoning and thoughtTurns is 0.

#turns

#parts[]

#excludedAttachments

The number of attachments this turn had that are not in the export, because scrapeImages or scrapeAttachments excluded them.

It exists because removing them outright left messages that were nothing but a heading and a token count, which reads as a bug. If you are treating an export as a dataset, check this field: a turn with excludedAttachments > 0 is incomplete, and you cannot tell that from parts alone.

#JSONL

One JSON object per line. No wrapper array, no trailing comma, \n-terminated.

{"messages": [{"role": "system", "content": "…"}, {"role": "user", "content": "…"}, {"role": "assistant", "content": "…"}]}

system appears only if the conversation had a system instruction. Note that it is included whenever one exists; the includeSystemInstruction setting strips it upstream, so turning that off removes it here too.

Line-count caveat: the number of lines is not the number of conversations you selected. A conversation with no assistant reply produces no line at all. The export's own summary reports what actually landed in the file.

#jsonlIncludeReasoning

Off by default. When off, reasoning turns are skipped entirely. When on, each becomes an assistant message and (because it is adjacent to the real answer, which is also assistant) is merged into it with a blank line between.

#CSV

One row per turn. UTF-8 with a BOM, CRLF line endings, every field quoted.

Formula-injection guard: any cell whose text begins =, +, -, @, tab or CR is prefixed with a single apostrophe. Strip a leading ' if you are parsing programmatically and the original value matters.

#Library manifest

A whole-library export to a per-conversation format writes two extra files at the root of each archive.

index.md is a human-readable listing linking to every conversation, plus any failures.

manifest.json:

{
  "exporter": "ai-studio-exporter",
  "exportedAt": "2026-07-31T06:40:38.000Z",
  "format": "markdown",
  "volume": 1,
  "volumeCount": 2,
  "library": [
    {
      "title": "How does the RPC auth work?",
      "path": "how-does-the-rpc-auth-work/conversation.md",
      "modifiedAt": 1750000000000,
      "turnCount": 5,
      "volume": 1
    },
    {
      "title": "Design Review",
      "path": "design-review/conversation.md",
      "modifiedAt": 1749000000000,
      "turnCount": 210,
      "volume": 2
    }
  ],
  "conversations": [
    {
      "title": "How does the RPC auth work?",
      "path": "how-does-the-rpc-auth-work/conversation.md",
      "modifiedAt": 1750000000000,
      "turnCount": 5,
      "volume": 1
    }
  ],
  "failures": []
}

#Value glossary

#source

"api" is read from AI Studio's own backend. Full fidelity: token counts, timestamps, model configuration, system instruction, and attachment bytes.

"dom" is scraped from the rendered page, used when the API path is unavailable. Only what the page displays: no token counts, no model config, no system instruction, and attachments limited to what was rendered. If fields are missing across the board, check this first.

#isThought

Marks a model turn as reasoning rather than the answer. Reasoning is a sibling of the answer in the turn list, not a child of it, because that is how AI Studio stores it. A question typically produces two model turns in sequence, the first with isThought: true.

#Token counts

AI Studio's own numbers, passed through unmodified; they are not a re-count, and not comparable to a tokenizer you run yourself. A turn with no reported count is null, never 0.

When condenseAttachments folds attachment turns into a message, their token counts are summed into the surviving turn, so a conversation's total is unchanged by the setting.

#Attachment type

image or file, decided from the MIME type rather than the extension. This is what scrapeImages and scrapeAttachments filter on, and what decides whether Markdown and HTML embed the file inline or link to it.