Import data

Import large datasets efficiently from Amazon S3, Google Cloud Storage, or Azure Blob Storage into Pinecone serverless indexes using object storage.

Importing from object storage is the most efficient and cost-effective way to load large numbers of documents into an index.

To run through this guide in your browser, see the Bulk import colab notebook.

This feature is available on Standard and Enterprise plans.

​

Before you import

Before you can import data, ensure you have a serverless index, a storage integration, and data uploaded to an Amazon S3 bucket, Google Cloud Storage bucket, or Azure Blob Storage container.

Your uploaded data must be in JSON Lines (JSONL) format for a document index, or Parquet for a legacy vector index. If your source data isn’t already in that format, prepare it in that format before uploading.

​

1. Create an index

Create a serverless index for your data.

Be sure to create your index on a cloud that supports importing from the object storage you want to use:

	…to an AWS index	…to a GCP index	…to an Azure index
Import from AWS S3…	✅	❌	❌
Import from Google Cloud Storage…	✅	✅	✅
Import from Azure Blob Storage…	✅	✅	✅

​

2. Add a storage integration

To import from a public data source, a storage integration isn’t required. However, to import from a secure data source, you must create an integration to allow Pinecone access to data in your object storage. See the following guides:

• Integrate with Amazon S3
• Integrate with Google Cloud Storage
• Integrate with Azure Blob Storage

​

3. Prepare your data

1. Create a directory for each namespace

   In your Amazon S3 bucket, Google Cloud Storage bucket, or Azure Blob Storage container, create an import directory containing a subdirectory for each namespace you want to import into. The namespaces must not yet exist in your index.

   For example, to import data into the namespaces example_namespace1 and example_namespace2, your directory structure would look like this:

   <BUCKET_OR_CONTAINER_NAME>/
   --/<IMPORT_DIR>/
   ----/example_namespace1/
   ----/example_namespace2/
   

   To import into the default namespace, use a subdirectory called __default__. The default namespace must be empty.

2. Create a data file for each namespace

   For each namespace, create one or more files defining the data to import.

   To import into a namespace in an index with a document schema, use JSONL (.jsonl, or gzip-compressed .jsonl.gz) files. Each line is one document, identical in shape to a document you would pass to the Upsert documents operation:

   Field	JSON type	Description
   _id	string	Required. Unique identifier for each document within the namespace.
   Each schema field	Depends on the field’s type	Encode each schema-declared field by its type: a full-text string field as a JSON string; a dense_vector field as an array of floats matching the schema’s dimension; a sparse_vector field as {"indices": [...], "values": [...]}. A document doesn’t need to populate every declared field, but it must include _id and at least one schema field.
   Any other field	string, number, boolean, or array of strings	Optional. Stored and auto-indexed as filterable metadata. Field names can’t start with _ or $.

   For example, for a schema with a full-text body field and a dense_vector embedding field:

   {"_id": "doc1", "body": "Machine learning models are revolutionizing natural language processing", "embedding": [0.12, 0.34, 0.56], "category": "technology", "year": 2024}
   {"_id": "doc2", "body": "Vector databases enable fast similarity search across embeddings", "embedding": [0.91, 0.05, 0.44], "category": "technology", "year": 2023}
   

   For the full file format, per-field encoding, and directory layout, see Prepare document-schema files.

   To import into a legacy vector index, use Parquet files instead. See Import vectors.

3. Upload the files

   Upload the files into the relevant subdirectory.

   For example, if you have subdirectories for the namespaces example_namespace1 and example_namespace2 and upload 4 files into each, your directory structure would look as follows after the upload. Files use the .jsonl or .jsonl.gz extension for document indexes, or .parquet for legacy vector indexes:

   <BUCKET_OR_CONTAINER_NAME>/
   --/<IMPORT_DIR>/
   ----/example_namespace1/
   ------0.<EXT>
   ------1.<EXT>
   ------2.<EXT>
   ------3.<EXT>
   ----/example_namespace2/
   ------4.<EXT>
   ------5.<EXT>
   ------6.<EXT>
   ------7.<EXT>
   

​

4. Import data into an index

Review import limits before starting an import.

The same request imports JSONL files into a document index or Parquet files into a legacy vector index. You can’t import into an index whose schema declares embed or sparse_embed.

Use the Start import operation to start an asynchronous import from object storage into an index.

• For uri, specify the URI of the bucket and import directory containing the namespaces and files you want to import. For example:

  • Amazon S3: s3://BUCKET_NAME/IMPORT_DIR
  • Google Cloud Storage: gs://BUCKET_NAME/IMPORT_DIR
  • Azure Blob Storage: https://STORAGE_ACCOUNT.blob.core.windows.net/CONTAINER_NAME/IMPORT_DIR

• For integration_id, specify the Integration ID of the Amazon S3, Google Cloud Storage, or Azure Blob Storage integration you created. The ID is found on the Storage integrations page of the Pinecone console.

  An Integration ID isn’t needed to import from a public bucket.

• For error_mode, use continue or abort.

  • With abort, the operation stops if any documents fail to import.
  • With continue, the operation continues on error, but there isn’t any notification about which documents, if any, failed to import. To see how many were successfully imported, use the describe an import operation.

Python

from pinecone import Pinecone, ImportErrorMode

pc = Pinecone(api_key="YOUR_API_KEY")

# To get the unique host for an index, 
# see https://docs.pinecone.io/guides/indexes/target-an-index
index = pc.index(host="INDEX_HOST")
root = "s3://example_bucket/import"

index.start_import(
    uri=root,
    integration_id="a12b3d4c-47d2-492c-a97a-dd98c8dbefde", # Optional for public buckets
    error_mode=ImportErrorMode.CONTINUE # or ImportErrorMode.ABORT
)

The response contains an id that you can use to check the status of the import:

Response

{
   "id": "101"
}

Once all the data is loaded, the index builder indexes the data, which usually takes at least 10 minutes. During this indexing process, the expected job status is InProgress, but 100.0 percent complete. Once all the imported records are indexed and fully available for querying, the import operation is set to Completed. If you cancel the import before it finishes, the status changes to Cancelled.

You can start a new import using the Pinecone console. Find the index you want to import into, and click the ellipsis (…) menu > Import data.

​

5. Track import progress

The amount of time required for an import depends on various factors, including:

• The number of documents to import
• The number of namespaces to import, and the number of documents in each
• The total size (in bytes) of the import

To track an import’s progress, check its status bar in the Pinecone console or use the Describe an import operation with the import ID:

Python

from pinecone import Pinecone

pc = Pinecone(api_key="YOUR_API_KEY")

# To get the unique host for an index, 
# see https://docs.pinecone.io/guides/indexes/target-an-index
index = pc.index(host="INDEX_HOST")

index.describe_import(id="101")

The response contains the import details, including the import status, percent_complete, and records_imported:

Response

{
  "id": "101",
  "uri": "s3://example_bucket/import",
  "status": "InProgress",
  "created_at": "2024-08-19T20:49:00.754Z",
  "finished_at": "2024-08-19T20:49:00.754Z",
  "percent_complete": 42.2,
  "records_imported": 1000000
}

If the import fails, the response contains an error field with the reason for the failure. See the Troubleshooting section for more information.

Response

{
  "id": "102",
  "uri": "s3://example_bucket/import",
  "status": "Failed",
  "percent_complete": 0.0,
  "records_imported": 0,
  "created_at": "2025-08-21T11:29:47.886797+00:00",
  "error": "User error: The namespace \"namespace1\" already exists. Imports are only allowed into nonexistent namespaces.",
  "finished_at": "2025-08-21T11:30:05.506423+00:00"
}

​

Prepare document-schema files (JSONL)

For indexes with a document schema, import files are JSON Lines (.jsonl), optionally gzip-compressed (.jsonl.gz), instead of Parquet. Each line is one document, identical in shape to a document you would pass to the Upsert documents operation, and validated through the same code path, so a document that upserts cleanly imports cleanly.

​

File format

For example, consider an index whose schema declares one full-text field and one dense vector field:

{
  "schema": {
    "fields": {
      "body": {
        "type": "string",
        "full_text_search": { "language": "en" }
      },
      "embedding": {
        "type": "dense_vector",
        "dimension": 3,
        "metric": "cosine"
      }
    }
  }
}

Each line supplies _id, the schema-declared fields, and any additional fields you want stored as filterable metadata (category and year here):

{"_id": "doc1", "body": "Machine learning models are revolutionizing natural language processing", "embedding": [0.12, 0.34, 0.56], "category": "technology", "year": 2024}
{"_id": "doc2", "body": "Vector databases enable fast similarity search across embeddings", "embedding": [0.91, 0.05, 0.44], "category": "technology", "year": 2023}

Each document follows these rules:

• _id is required in every document: a non-empty string, unique within the namespace.
• Each document must include at least one schema-declared field; it doesn’t need to populate every declared field.
• Searchable fields declared in the schema are encoded by type:
  • A dense_vector field (embedding above) is a JSON array of floats whose length equals the schema’s declared dimension.
  • A sparse_vector field is an object {"indices": [...], "values": [...]}.
  • A full-text string field (body above) is a JSON string.
• Any other field (category and year above) is stored on the document and auto-indexed as filterable metadata: a string, number, boolean, or array of strings. An array of numbers in an undeclared field is rejected rather than stored as metadata.
• Field names must not start with _ (reserved for system fields like _id) or $ (reserved for filter operators).

The same per-document size limits as upsert apply.

​

Directory layout

Lay files out by namespace, as described in Prepare your data: under a common dataset prefix in your bucket or container, create one subdirectory per namespace and place that namespace’s .jsonl / .jsonl.gz files inside it.

<BUCKET_OR_CONTAINER_NAME>/
└── <DATASET_PREFIX>/
    ├── namespace1/
    │   ├── 0.jsonl
    │   └── 1.jsonl.gz
    └── namespace2/
        └── 0.jsonl

• Each namespace must not already exist in the index — imports only create new namespaces.
• A namespace subdirectory can mix plain .jsonl and gzip-compressed .jsonl.gz files.

​

Prepare a file

Generate a schema-conformant .jsonl (or .jsonl.gz) file from your documents, then upload it to a namespace subdirectory in object storage.

Python

import gzip
import json

documents = [
    {
        "_id": "doc1",
        "body": "Machine learning models are revolutionizing natural language processing",
        "embedding": [0.12, 0.34, 0.56],   # must match the schema dimension
        "category": "technology",
        "year": 2024,
    },
    {
        "_id": "doc2",
        "body": "Vector databases enable fast similarity search across embeddings",
        "embedding": [0.91, 0.05, 0.44],
        "category": "technology",
        "year": 2023,
    },
]

# Plain JSONL: one compact JSON document per line.
with open("0.jsonl", "w") as f:
    for doc in documents:
        f.write(json.dumps(doc) + "\n")

# Or gzip-compressed (.jsonl.gz) to cut storage and transfer.
with gzip.open("0.jsonl.gz", "wt") as f:
    for doc in documents:
        f.write(json.dumps(doc) + "\n")

Upload the file to the relevant namespace subdirectory in your bucket or container, e.g. <DATASET_PREFIX>/namespace1/0.jsonl.

​

Convert Parquet files to JSONL

If your data is already stored as Parquet in the vector import format (id, values, sparse_values, metadata columns), convert each file to JSONL for the document schema. Map id to _id, the vector columns to your schema’s field names, and spread the metadata JSON string into top-level fields:

Python

import gzip
import json

import pyarrow.parquet as pq

# Your schema's field names
DENSE_FIELD = "embedding"          # dense_vector field
SPARSE_FIELD = "sparse_embedding"  # sparse_vector field

table = pq.read_table("0.parquet")

with gzip.open("0.jsonl.gz", "wt") as f:
    for row in table.to_pylist():
        doc = {"_id": row["id"]}
        if row.get("values") is not None:
            doc[DENSE_FIELD] = row["values"]
        if row.get("sparse_values") is not None:
            doc[SPARSE_FIELD] = {
                "indices": row["sparse_values"]["indices"],
                "values": row["sparse_values"]["values"],
            }
        if row.get("metadata"):
            # Metadata keys become top-level fields. Keys that match a
            # schema-declared field (e.g. a full-text "body" field) are
            # validated against the schema; the rest are auto-indexed metadata.
            doc.update(json.loads(row["metadata"]))
        f.write(json.dumps(doc) + "\n")

​

Manage imports

​

List imports

Use the List imports operation to list all of the recent and ongoing imports. By default, the operation returns up to 100 imports per page. If the limit parameter is passed, the operation returns up to that number of imports per page instead. For example, if limit=3, up to 3 imports are returned per page. Whenever there are additional imports to return, the response includes a pagination_token for fetching the next page of imports.

Python SDK

When using the Python SDK, list_import paginates automatically.

Python

from pinecone import Pinecone, ImportErrorMode

pc = Pinecone(api_key="YOUR_API_KEY")

# To get the unique host for an index, 
# see https://docs.pinecone.io/guides/indexes/target-an-index
index = pc.index(host="INDEX_HOST")

# List using a generator that handles pagination
for i in index.list_imports():
    print(f"id: {i.id} status: {i.status}")

# List using a generator that fetches all results at once
operations = list(index.list_imports())
print(operations)

Response

{
  "data": [
    {
      "id": "1",
      "uri": "s3://BUCKET_NAME/PATH/TO/DIR",
      "status": "Pending",
      "started_at": "2024-08-19T20:49:00.754Z",
      "finished_at": "2024-08-19T20:49:00.754Z",
      "percent_complete": 42.2,
      "records_imported": 1000000
    }
  ],
  "pagination": {
    "next": "Tm90aGluZyB0byBzZWUgaGVyZQo="
  }
}

You can view the list of imports for an index in the Pinecone console. Select the index and navigate to the Imports tab.

Other SDKs

When using the Node.js SDK, Java SDK, Go SDK, or REST API to list recent and ongoing imports, you must manually fetch each page of results. To view the next page of results, include the paginationToken provided in the response.

JavaScript

import { Pinecone } from '@pinecone-database/pinecone';

const pc = new Pinecone({ apiKey: 'YOUR_API_KEY' });

// To get the unique host for an index, 
// see https://docs.pinecone.io/guides/indexes/target-an-index
const index = pc.index("INDEX_NAME", "INDEX_HOST")

const results = await index.listImports({ limit: 10, paginationToken: 'Tm90aGluZyB0byBzZWUgaGVyZQo' });
console.log(results);

​

Cancel an import

The Cancel an import operation cancels an import if it isn’t yet finished. It has no effect if the import is already complete.

Python

from pinecone import Pinecone

pc = Pinecone(api_key="YOUR_API_KEY")

# To get the unique host for an index, 
# see https://docs.pinecone.io/guides/indexes/target-an-index
index = pc.index(host="INDEX_HOST")

index.cancel_import(id="101")

Response

{}

You can cancel your import using the Pinecone console. To cancel an ongoing import, select the index you are importing into and navigate to the Imports tab. Then, click the ellipsis (…) menu > Cancel.

​

Import limits

If your import exceeds these limits, you’ll get an error specifying the limit exceeded. See Troubleshooting for details.

Metric	Limit
Max namespaces per import	10,000
Max total input data size (on-demand indexes)	1 TB
Max total input data size (DRN indexes)	Unlimited
Max files per import	100,000
Max size per file	10 GB

The total input data size limit doesn’t apply to indexes with dedicated read nodes.

Bulk import supports indexes without a schema definition (Parquet files) and indexes with document schemas (JSONL files). Semantic-text (auto-embedded) fields aren’t yet supported in document schemas.

Also:

• You can’t import data from an AWS S3 bucket into a Pinecone index hosted on GCP or Azure.
• You can’t import data from S3 Express One Zone storage.
• You can’t import data into an existing namespace.
• When importing data into the __default__ namespace of an index, the default namespace must be empty.
• Each import takes at least 10 minutes to complete.
• You can’t import into an index whose schema declares embed or sparse_embed. To add text that Pinecone embeds, use upsert.

​

Troubleshooting

When an import fails, you’ll see an error message with the reason for the failure in the Pinecone console or in the response to the describe an import operation.

Namespace already exists

You can’t import data into an existing namespace. If your import directory structure contains a folder with the name of an existing namespace in your index, the import will fail with the following error:

User error: The namespace "example-namespace" already exists. Imports are only allowed into nonexistent namespaces.

To fix this, rename the folder to use a namespace name that doesn’t yet exist.

No namespace found

In object storage, your directory structure must be as follows:

example_bucket/
--/imports/
----/example_namespace1/
------0.parquet
------1.parquet
------2.parquet
------3.parquet
----/example_namespace2/
------4.parquet
------5.parquet
------6.parquet
------7.parquet

If a Parquet file isn’t nested under a namespace subdirectory, the import will fail with the following error:

User error: "test-import/0.parquet": No namespace detected. Each file should be nested under a subdirectory of the URI prefix. This indicates which namespace it should be imported into.

To fix this, move the Parquet file to a namespace subdirectory.

Parquet files not found

Each namespace subdirectory must contain Parquet files with data to import. If a namespace subdirectory doesn’t include Parquet files, the import will fail with the following error:

User error: No Parquet files found under "gs://example_bucket/imports". Files must be stored with the specified bucket prefix.

To fix this, add Parquet files to the namespace subdirectory.

Namespace not created (empty subdirectory)

For document-schema (JSONL) imports, an empty namespace subdirectory behaves differently than for Parquet. A namespace subdirectory that contains no .jsonl or .jsonl.gz files is silently skipped: the import doesn’t create that namespace and doesn’t raise an error for it. If a namespace is missing after an import that otherwise succeeded, confirm its subdirectory actually contains importable files.

If no importable files are found anywhere under the dataset prefix, the import fails with the same No Parquet files found error shown above, which says “Parquet files” even for JSONL imports.

Invalid import URI

In your start import request, the import uri must specify only the bucket and import directory containing the namespaces and Parquet files you want to import. If the uri also contains a namespaces directory or a Parquet filename, the import will fail with the following error:

User error: "test-import/0.parquet": It looks like you specified a complete path to a parquet file as the URI prefix to import from. Note that the URI prefix should give an ancestor directory with subdirectories to specify each namespace to import into. See https://docs.pinecone.io/guides/data/understanding-imports#directory-structure.

To fix this, remove the namespaces directory or Parquet filename from the uri.

Invalid Parquet files

When a Parquet file isn’t formatted correctly, the import will fail with a message like one of the following:

File schema errors

Missing required column "{0}"
Unsupported column "{0}"

File corruption errors

Parquet footer could not be parsed. Are you sure this is valid parquet?

Type errors

The expected data type for column "{column}" is "{expected}", but got "{given}"
The expected data type for metadata is a JSON encoded string in UTF-8 format, but got "{given}"

These errors are returned for both continue and abort error modes.

To fix these errors, check the specific error message and follow the instructions in the Prepare your data section.

Invalid JSONL files or documents

For document-schema indexes, import files are JSONL: each line is one JSON document, validated against the index’s schema. A line fails if it isn’t valid JSON, or if the document doesn’t conform to the schema. Each per-document error identifies the file name, row number, and document _id:

error reading record (file "0.jsonl", row 2, id "doc1"): {reason}

Common reasons include:

Malformed JSON

invalid JSON document: {details}

Empty _id

Vector ID must not be empty

Dense-vector dimension mismatch

Vector dimension {actual} does not match the dimension of the index {expected}

Reserved field name

Document with id '{id}': Field name '{name}' cannot start with '_' (reserved for internal use)

Wrong field type

Document with id '{id}': full text search field '{field}' must be a string

A document that omits _id entirely fails JSON parsing and appears as a malformed-JSON error (missing field _id).

With errorMode.onError set to continue (the default), invalid documents are skipped and the rest import; with abort, the import stops on the first invalid document. If every document in a namespace is skipped, the import fails with No vectors added, all rows were skipped for namespace: {namespace}.

To fix these errors, validate your documents against the file-format rules before importing.

Invalid records

When the error_mode is abort and a file contains invalid records, the import will stop processing on the first invalid record and return an error message identifying the file name and row:

User error: error reading record (file "/0.parquet", row 0):

This will be followed by an error message identifying the specific issue. For example:

Missing values

missing required values in column "{column}"

Invalid metadata

Failed to parse metadata: {msg}

Invalid vectors

Upserting dense vectors is not supported for indexes that store only sparse vectors

When the error_mode is continue, the import will skip individual invalid records. However, if all records are invalid and skipped (for example, the vector type in the file doesn’t match the vector type of the index), the import will fail with a general message:

User error: No vectors added, all rows were skipped for namespace: example-namespace

To fix these errors, check the specific error message and follow the instructions in the Prepare your data section.

Duplicate records

Bulk import deduplicates by record ID, not by vector value. If an import contains more than one record with the same ID in the same namespace, only one is kept and the rest are counted as skipped. Which record is kept isn’t deterministic. Records that share identical vector values but have different IDs are all imported.

This applies to both JSONL imports (document indexes) and Parquet imports (legacy vector indexes). To control which record wins, deduplicate your source data by ID before importing.

Example scenario: If your import file contains:

id | values
---|---------
1  | [0.1, 0.2, 0.3]
1  | [0.4, 0.5, 0.6]  ← Same ID as the first record; only one of the two is kept
2  | [0.7, 0.8, 0.9]

Two records are imported: 2, and one of the two records with ID 1.

Import exceeds maximum data size for on-demand

On-demand indexes have a maximum total input data size of 1 TB per import. If your import exceeds this limit, it will fail with the following error:

Import ({size} GB) exceeds the maximum input data size of 1000 GB for on-demand. Consider using Dedicated Read Nodes (DRN) for larger index sizes, or contact support for your use-case.

To fix this, either reduce the total size of your import to under 1 TB, use an index with dedicated read nodes (which have no total data size limit for imports), or contact support.

For .jsonl.gz files, size is measured as an estimated uncompressed size of 10× the compressed file, so gzip-compressed files count roughly 10× their on-disk size against this limit.

​

See also

• Integrate with Amazon S3
• Integrate with Google Cloud Storage
• Integrate with Azure Blob Storage
• Pinecone’s pricing