LOAD CSV

Overview

LOAD CSV statement has two forms:

• Dump CSV form reads a CSV file and returns each row as a map. Use this to preview a file.
• Import from CSV form writes each CSV row straight into the current graph as a node or edge. Use this to load data. Large files are streamed in batches.

Both forms read a local filesystem path, a file:// URI, or an http(s):// URL. Other remote schemes (s3://, ftp://, ...) are not supported.

Syntax
<load csv statement> ::= 
  "LOAD CSV FROM" <single-quoted file source> { <dump csv> | <import csv> }

<dump csv> ::=
  [ "AS" <row variable> ]
  [ "WITH HEADER" [ "DELIMITER" <character> [ "QUOTE" <character> ] ] ]

<import csv> ::=
  "WITH HEADER" [ "DELIMITER" <character> [ "QUOTE" <character> ] ]
  "INTO" <label name>
  [ "EDGE FROM" <label name> "(" <column> ")" "TO" <label name> "(" <column> ")" ]
  [ "MAPPING" "(" <property mapping> { "," ... } ")" ]

<property mapping> ::= <property> ":" <column> [ "AS" <type> ]

Details

• <row variable> defaults to row when AS is omitted.
• With WITH HEADER, the first row is consumed as column names and each subsequent row is keyed by those names. Without WITH HEADER, each row is keyed positionally: col0, col1, ... .
• DELIMITER accepts a single character (e.g. ',', ';', '\t'). QUOTE accepts the double quote (") as the field-quote character.
• About the file source:
  • Path resolution: absolute paths (/data/users.csv) and file:/// URIs are used as-is. Relative paths resolve against the server process's working directory, the directory the database was launched from, not the .gdb data folder. Prefer absolute paths in production.
  • HTTP(S) sources are streamed directly into the CSV reader (no temp file). Connect, TLS-handshake, and response-header waits are bounded (30 s, 30 s, 60 s); the body itself streams without a total deadline and is cancelled if the query is cancelled. Non-200 responses fail the statement with the HTTP status.

Example CSVs

user.csv
_id,name,balance,risk_score,is_mule
USR_000000,David Brown,16625.7,0.86,true
USR_000001,James Johnson,62870.63,0.76,true
USR_000002,Jane Brown,17065.87,0.85,true

transfers.csv
_from,_to,amount,is_suspicious,timestamp
USR_000000,USR_000002,3384.45,false,2024-01-27T19:19:00Z
USR_000001,USR_000000,4353.22,true,2024-01-25T06:18:00Z
USR_000000,USR_000001,2482.13,false,2024-01-30T15:27:00Z
USR_000001,USR_000002,4050.61,false,2024-01-28T01:02:00

Dump CSV to Inspect Rows

Read a CSV file with no header. Each row is returned as a map keyed positionally (col0, col1, ...) under a single column named row (the default variable):

GQL
LOAD CSV FROM 'data/users.csv'

Result:

JSON
[
  {
    "row": {
      "col2": "balance",
      "col3": "risk_score",
      "col4": "is_mule",
      "col0": "_id",
      "col1": "name"
    }
  },
  {
    "row": {
      "col0": "USR_000000",
      "col1": "David Brown",
      "col2": "16625.7",
      "col3": "0.86",
      "col4": "true"
    }
  },
  {
    "row": {
      "col0": "USR_000001",
      "col1": "James Johnson",
      "col2": "62870.63",
      "col3": "0.76",
      "col4": "true"
    }
  },
  {
    "row": {
      "col3": "0.85",
      "col4": "true",
      "col0": "USR_000002",
      "col1": "Jane Brown",
      "col2": "17065.87"
    }
  }
]

Read a CSV file with a header so the map is keyed by column name:

GQL
LOAD CSV FROM 'data/users.csv' WITH HEADER

Result:

JSON
[
  {
    "row": {
      "_id": "USR_000000",
      "name": "David Brown",
      "balance": "16625.7",
      "risk_score": "0.86",
      "is_mule": "true"
    }
  },
  {
    "row": {
      "_id": "USR_000001",
      "name": "James Johnson",
      "balance": "62870.63",
      "risk_score": "0.76",
      "is_mule": "true"
    }
  },
  {
    "row": {
      "_id": "USR_000002",
      "name": "Jane Brown",
      "balance": "17065.87",
      "risk_score": "0.85",
      "is_mule": "true"
    }
  }
]

Bind the per-row map to a different column name with AS:

GQL
LOAD CSV FROM '/data/users.csv' AS u

Read a tab-separated file:

GQL
LOAD CSV FROM 'data/users.tsv' AS u WITH HEADER DELIMITER '\t'

Specify the field-quote character with QUOTE. The double quote (") is the supported quoting character:

GQL
LOAD CSV FROM 'data/users.csv' AS u WITH HEADER DELIMITER ',' QUOTE '"'

Read directly from an HTTP(S) URL — the file is streamed, not downloaded to disk first:

GQL
LOAD CSV FROM 'https://example.com/data/users.csv' WITH HEADER

The result set has one column (row / u / whatever name AS specified). Each row in that column is a MAP — access fields client-side after fetching, e.g. row['name'] in your driver.

The rows are returned to the client only, they cannot be passed to a following RETURN, INSERT, or MATCH in the same query.

Import from CSV

Use INTO <label> to write each CSV row as a node (or as an edge with EDGE FROM ... TO ...) in a single statement, no chained INSERT needed. Returns a one-row summary with the count of imported (and, for edges, skipped) records.

Node import

Without MAPPING, every CSV column becomes a property of the same name on the new node. WITH HEADER is required (the import needs column names to map by). All values default to STRING.

GQL
LOAD CSV FROM 'data/users.csv' WITH HEADER INTO User

Add a MAPPING (...) clause to rename columns, select a subset, and apply type coercion. Include _id in the mapping if you plan to reference these nodes from a later edge import.

GQL
LOAD CSV FROM 'data/users.csv' WITH HEADER
INTO Person MAPPING (
  _id: '_id',
  name: 'name',
  balance: 'balance' AS FLOAT,
  riskScore: 'risk_score' AS FLOAT,
  isMule: 'is_mule' AS BOOL
)

Edge import

For edges, name the endpoint labels and the CSV columns that supply each endpoint's _id. Rows whose endpoint _id is missing in the graph are skipped and reported in the summary instead of erroring the whole import. Requires the graph to have edge _id enabled (default for new graphs).

GQL
LOAD CSV FROM 'data/transfers.csv' WITH HEADER
INTO TRANSFERS EDGE FROM Person('_from') TO Person('_to')
MAPPING (
  isSuspicious: 'is_suspicious' AS BOOL,
  timestamp: 'timestamp' AS TIMESTAMP
)  

Supported Type Coercions

Notes:

• Empty CSV cells produce a NULL (the property is omitted from the row) rather than an empty-string value.
• Coercion failures error out with the row number and the offending column name.

Limitations

• Every cell is read as a STRING in the dump form; downstream client code is responsible for any type conversion. The inline-import form does its own coercion via AS <TYPE> in the MAPPING clause.