Export format

An export is one archive of everything your tenant holds: your documents and their text, the passages and pages read from them, your entities, relationships and memory, your schema and its history, and your audit log. Your administrators request it from the console's Export your data page. It is ready once the archive is written, and it can be downloaded for seven days. While a deletion of your tenant can still be cancelled, an export made before it can still be downloaded.

This page is generated from a snapshot of what the export code writes, and a test fails if the two differ. Every field of every file is listed below.

The archive

The archive is a POSIX tar file (not compressed). Its first member is manifest.json; the other files follow in the order listed below, each exactly once, even when it holds no records.

Every .jsonl file is JSON Lines: one JSON object per line, in UTF-8, each line ending with a line feed. Each object's keys are in the canonical order of RFC 8785 (JSON Canonicalization Scheme), so the same data always produces the same bytes. manifest.json is one canonical JSON object. anchor.json is one canonical JSON object, or null when the audit log has no anchor yet.

All files are read from one consistent snapshot of your tenant's data, taken when the export ran: a record changed after that moment is not in the export, and every record in it is as it was at that moment.

Checking an export

manifest.json lists every file with its number of records, its size in bytes and its SHA-256 digest. To check an archive, extract it and compare each file's SHA-256 (for example with sha256sum) and size with the manifest's. A file that is missing, extra, or differs is a damaged archive: request a new export.

audit.jsonl holds your audit log up to the manifest's snapshotPosition, in position order, so it can be read as the history of every change in the export. anchor.json holds the newest anchor of that log, or null when the log has none yet.

How values are written

Each line of a .jsonl file is one stored record with every one of its fields, so a field can be null only where the tables below say it can. The JSON column says how each value is written:

JSON Written as
string A JSON string.
integer, number A JSON number.
boolean true or false.
timestamp A string in ISO 8601 with a UTC offset, such as 2026-11-14T09:30:00.123456+00:00.
date A string such as 2026-11-14.
json The JSON value itself (an object or an array), not a string holding it.
array A JSON array.
hex A string of \x followed by the bytes in lowercase hexadecimal.
base64 A string of standard base64 (RFC 4648), without line breaks.
base64url A string of URL-safe base64 (RFC 4648 §5), without padding.

The Stored as column names the database type the value comes from, for readers who load an export into their own database.

manifest.json

Format kg-export/1. The first file of the archive; it names every other file.

Fields of the manifest
FieldMeaning
formatThe format's name and version, kg-export/1.
tenantYour tenant's identifier.
exportIdThe export's identifier, as the console shows it.
snapshotAtWhen the snapshot the export was read from was taken (ISO 8601, UTC).
snapshotPositionThe audit position the snapshot holds the log through: audit.jsonl ends at this position (0 when the log is empty).
filesOne entry for every other file, in archive order.
Fields of each entry in files
FieldMeaning
nameThe file's name in the archive.
recordsHow many records the file holds: its lines, or for anchor.json 1 (0 when it holds null).
bytesThe file's size in bytes.
sha256The SHA-256 digest of the file's bytes, in lowercase hexadecimal.

entity-types.jsonl

Your entity types, one per line, ordered by name.

Fields of entity-types.jsonl
FieldJSONCan be nullStored asMeaning
storage_idstringnotextThe type definition's internal identifier.
tenant_idstringnotextYour tenant's identifier; the same on every record.
namestringnotextThe type's name.
identifying_propertystringnotextThe property whose value identifies an entity of this type.
properties_jsonjsonnotextThe type's property definitions, with each property's status and constraints.
created_seqintegernobigintA number that increases with each type created, giving creation order.
statusstringnotextWhether the type is active or deprecated.
changed_revisionintegernobigintThe schema revision of the type's last change (0 for a type older than schema history).

relationship-types.jsonl

Your relationship types, one per line, ordered by name.

Fields of relationship-types.jsonl
FieldJSONCan be nullStored asMeaning
storage_idstringnotextThe type definition's internal identifier.
tenant_idstringnotextYour tenant's identifier; the same on every record.
namestringnotextThe type's name.
allowed_source_typesarraynoARRAYThe entity types a relationship of this type may start from.
allowed_target_typesarraynoARRAYThe entity types a relationship of this type may point to.
properties_jsonjsonnotextThe type's property definitions, with each property's status and constraints.
created_seqintegernobigintA number that increases with each type created, giving creation order.
statusstringnotextWhether the type is active or deprecated.
changed_revisionintegernobigintThe schema revision of the type's last change (0 for a type older than schema history).

schema-history.jsonl

Every applied schema change, one per line, in revision order.

Fields of schema-history.jsonl
FieldJSONCan be nullStored asMeaning
tenant_idstringnotextYour tenant's identifier; the same on every record.
revisionintegernobigintThe schema revision this change produced, counting from 1.
committed_attimestampnotimestamp with time zoneWhen the change was applied.
principal_idstringnotextThe identity that made the change.
actor_kindstringnotextWho made the change: a human or an agent.
agent_idstringyestextThe agent's identifier when an agent made the change; null otherwise.
request_idstringnotextThe identifier of the request that made the change.
operations_jsonjsonnotextThe change's operations in order, each with the element's definition before and after (for a very large element, the changed fields and a digest).

entities.jsonl

Every entity, one per line, ordered by identifier. Documents, passages and memory entries are entities too, so they appear here as well as in their own files.

Fields of entities.jsonl
FieldJSONCan be nullStored asMeaning
idstringnotextThe entity's identifier.
tenant_idstringnotextYour tenant's identifier; the same on every record.
typestringnotextThe entity's type.
identifying_valuestringyestextThe value of its type's identifying property, when it has one.
properties_jsonjsonnotextThe entity's properties.
created_seqintegernobigintA number that increases with each entity created, giving creation order.
identifying_digesthexyesbyteaThe SHA-256 of the identifying value, kept for lookups; null when there is no identifying value.
identifying_normalizedstringyestextThe identifying value with whitespace collapsed, compatibility characters folded (NFKC) and ASCII letters lower-cased, kept for duplicate suggestions; null when there is no identifying value, and briefly for entities stored before it existed.

relationships.jsonl

Every relationship, one per line, ordered by identifier.

Fields of relationships.jsonl
FieldJSONCan be nullStored asMeaning
idstringnotextThe relationship's identifier.
tenant_idstringnotextYour tenant's identifier; the same on every record.
typestringnotextThe relationship's type.
source_idstringnotextThe entity it starts from.
target_idstringnotextThe entity it points to.
properties_jsonjsonnotextThe relationship's properties.
created_seqintegernobigintA number that increases with each relationship created, giving creation order.

extraction-receipts.jsonl

A receipt for every committed extraction, one per line: which reviewed candidates became entities and relationships. Extraction is off during the trial, so this file is normally empty.

Fields of extraction-receipts.jsonl
FieldJSONCan be nullStored asMeaning
receipt_idstringnotextThe receipt's identifier.
preview_idstringnotextThe extraction preview the commit approved.
tenant_idstringnotextYour tenant's identifier; the same on every record.
actorstringnotextThe actor the preview was issued to.
owner_keystringnotextThe key of the caller who owned the preview; only that caller could commit it.
principal_idstringnotextThe identity that committed it.
actor_kindstringnotextWho committed it: a human or an agent.
approval_hashstringnotextA digest of the approval, used to recognise a repeated commit.
review_hashstringnotextA digest of the reviewed preview.
selection_hashstringnotextA digest of the selected candidates.
receipt_jsonjsonnotextThe receipt: when it was committed, the source's digest, the model, and each change with its candidate, operation, record and evidence spans. It holds no source text.

documents.jsonl

Every document, one per line, ordered by identifier.

Fields of documents.jsonl
FieldJSONCan be nullStored asMeaning
tenant_idstringnotextYour tenant's identifier; the same on every record.
document_idstringnotextThe document's identifier; it is part of every passage identifier.
external_idstringyestextYour own key for the document, if you gave one when uploading.
entity_idstringnotextThe Document entity that represents it in the graph.
current_revisionintegeryesintegerThe revision in use; null until the first upload is committed.
next_revisionintegernointegerThe number the next upload will get.
created_attimestampnotimestamp with time zoneWhen the document was first uploaded.
updated_attimestampnotimestamp with time zoneWhen it last changed.

revisions.jsonl

Every retained revision of every document, with its text, one per line, ordered by document and revision.

Fields of revisions.jsonl
FieldJSONCan be nullStored asMeaning
tenant_idstringnotextYour tenant's identifier; the same on every record.
document_idstringnotextThe document.
revisionintegernointegerThe revision's number, counting from 1.
statestringnotextWhether the revision is pending, current, superseded or discarded.
media_typestringnotextThe uploaded file's media type.
original_digeststringnotextThe SHA-256 of the file as uploaded, in lowercase hexadecimal.
original_bytesintegernointegerThe size of the file as uploaded, in bytes.
normalized_textstringyestextThe revision's text after normalization; passages are slices of it. Null for a discarded revision.
normalized_digeststringnotextThe SHA-256 of the normalized text, in lowercase hexadecimal.
code_point_countintegernointegerThe normalized text's length in Unicode code points.
normalization_ruleintegernointegerThe version of the normalization rules that produced the text.
splitting_ruleintegernointegerThe version of the rules that split it into passages.
passage_countintegeryesintegerHow many passages the revision has; null until it is committed.
uploaded_bystringnotextThe identity that uploaded it.
actor_kindstringnotextWho uploaded it: a human or an agent.
agent_idstringyestextThe agent's identifier when an agent uploaded it; null otherwise.
created_attimestampnotimestamp with time zoneWhen it was uploaded.
committed_attimestampyestimestamp with time zoneWhen it became current; null if it never did.
superseded_attimestampyestimestamp with time zoneWhen a newer revision replaced it; null while it is current.
page_totalintegeryesintegerFor a PDF, its number of pages; null otherwise.
images_skippedintegeryesintegerHow many images in the file were not read; null for a format that has none.
completebooleannobooleanFalse while any page of a PDF failed to be read.
no_textbooleannobooleanTrue for a PDF whose pages were all blank, which has no text and no passages.

passages.jsonl

Every passage of every committed revision, one per line, ordered by document, revision and position. A passage's text is not repeated: it is the slice of its revision's normalized_text from span_start to span_end.

Fields of passages.jsonl
FieldJSONCan be nullStored asMeaning
tenant_idstringnotextYour tenant's identifier; the same on every record.
document_idstringnotextThe document.
revisionintegernointegerThe revision.
ordinalintegernointegerThe passage's position in the revision.
span_startintegernointegerWhere the passage starts in the revision's normalized text, in code points from 0.
span_endintegernointegerWhere it ends, in code points (exclusive).
heading_pathstringnotextThe headings the passage sits under, in full.
entity_idstringyestextThe Passage entity while the revision is current; null after it is superseded.
pageintegeryesintegerFor a PDF, the page the passage starts on; null otherwise.

originals.jsonl

Each uploaded PDF or Word file exactly as it was uploaded, one per line, ordered by document and revision. Its media type, SHA-256 and size are its revision's media_type, original_digest and original_bytes.

Fields of originals.jsonl
FieldJSONCan be nullStored asMeaning
document_idstringnotextThe document.
revisionintegernointegerThe revision.
content_base64base64nobyteaThe file's bytes.

pages.jsonl

Every page of every PDF revision and what reading it produced, one per line, ordered by document, revision and page.

Fields of pages.jsonl
FieldJSONCan be nullStored asMeaning
tenant_idstringnotextYour tenant's identifier; the same on every record.
document_idstringnotextThe document.
revisionintegernointegerThe revision.
pageintegernointegerThe page's number, counting from 1.
labelstringyestextThe page's label in the PDF, if it has one.
statestringnotextWhether the page is pending, read, empty or failed.
reasonstringyestextWhy the page failed; null unless it failed.
textstringyestextThe page's text as read, in Markdown.
modelstringyestextThe model that read the page.
settings_keystringyestextThe rendering settings the page was read with.
instruction_versionintegeryesintegerThe version of the reading instruction used.
tokensintegernobigintThe tokens the model provider reported for reading the page, input and output together.
reusedbooleannobooleanTrue when the page was copied from an earlier upload of the same file, read the same way, instead of being read again.

audit.jsonl

Your audit log up to the manifest's snapshotPosition: every committed change, one per line, in position order.

Fields of audit.jsonl
FieldJSONCan be nullStored asMeaning
tenant_idstringnotextYour tenant's identifier; the same on every record.
positionintegernobigintThe entry's position in the log, consecutive from 1 in commit order.
committed_attimestampnotimestamp with time zoneWhen the change was committed.
principal_idstringnotextThe identity that made the change.
actor_kindstringnotextWho made the change: a human, an agent, or the service itself (system).
agent_idstringyestextThe agent's identifier when an agent made the change; null otherwise.
request_idstringnotextThe identifier of the request that made the change.
operationstringnotextWhat was done, such as create, update or delete.
record_kindstringnotextThe kind of record changed, such as entity, relationship or schema.
record_typestringyestextThe record's type, when it has one.
record_idstringyestextThe record's identifier, when it has one.
before_jsonjsonyestextThe record's complete state before the change; null for a creation.
after_jsonjsonyestextThe record's complete state after the change; null for a deletion.
refs_jsonjsonyestextOther records the change refers to, when there are any.
confirmationstringyestextFor an operation that needed confirmation, whether it was confirmed.
key_idstringnotextThe identifier of the key the entry's MAC was computed with.
macbase64urlnobyteaThe entry's message authentication code, which chains it to the previous entry, so the service can detect a changed or missing entry.

anchor.json

The newest anchor of your audit log at the snapshot: a record, kept outside your tenant's data, of a position in the log and the MAC of the entry there. It is null when the log has no anchor yet.

Fields of anchor.json
FieldMeaning
tenantYour tenant's identifier.
positionThe audit position the anchor was taken at.
entryMacThe MAC of the audit entry at that position, in URL-safe base64; it equals that entry's mac in audit.jsonl.
anchoredAtWhen the anchor was recorded (ISO 8601, UTC).