Turn your systems into training data
Connect a source once and the training file keeps itself current. Databases, SaaS, APIs, files and documents become validated JSONL datasets a fine-tuning job reads by ID.
The systems you already run
95 sources in eight kinds, from the catalogue: databases, SaaS, APIs, files, documents, messaging, analytics and industrial systems.
One file per source, kept current
Each source writes a JSONL file and refreshes it on the schedule you set, so the training set never goes stale.
Mapped onto the chat format
A template over the record's fields sets the system, user and assistant turns, with a filter for the records that qualify.
Validated before you spend GPU time
Every line is checked, and a rejected line comes back with its number and the reason.
Connect a source once. The training file keeps itself current.
Every source, filterable by kind
95 sources in eight kinds. Filter by kind or by what the system holds, search by name, and connect a source from its card. Each one writes a training file of its own.
Salesforce
SaaSCases, contacts and their resolutions
HubSpot
SaaSDeals, contacts and the notes on each
Zendesk
SaaSSupport tickets and the replies that closed them
Jira
SaaSIssues, comments and how each was resolved
SAP
SaaSMaster data, orders and their documents
Shopify
SaaSOrders, products and customer messages
PostgreSQL
DatabasesTables and rows as records
MySQL
DatabasesRows from any query you write
Snowflake
DatabasesWarehouse tables, filtered by a query
Databricks
DatabasesLakehouse tables and the results of a query
REST
APIsAny JSON endpoint, paged and mapped
GraphQL
APIsRecords from a query against your schema
Webhooks
APIsEvents pushed to you as they happen
CSV
FilesRows from a delimited file, with headers
Amazon S3
FilesObjects from a bucket, by prefix
Excel
FilesSheets and named ranges as rows
Google Drive
DocumentsDocs, sheets and slides, with their text
Microsoft 365
DocumentsDocuments, spreadsheets and site pages
Confluence
DocumentsWiki pages and the spaces holding them
Slack
MessagingChannel conversations and threads
Gmail
MessagingThreads, replies and their attachments
Google Analytics
AnalyticsSessions, events and traffic sources
Google Ads
AnalyticsCampaign spend, clicks and conversions
MQTT
IndustrialDevice telemetry by topic
What the connectors give you
Sources
- 95 sources in eight kinds
- SaaS, databases and warehouses, APIs, files and storage
- Documents, messaging, analytics and monitoring, industrial and IoT
- Filter by kind and by what a system holds
Mapping
- A template over the record's fields
- System, user and assistant turns
- A filter for the records that qualify
- Purposes: fine-tune, evaluation and retrieval
Refresh
- Hourly, daily, weekly or manual
- Every refresh writes a new validated file
- The file ID always names the latest ready version
- History kept newest first, with rejected lines and reasons
Control
- Credentials stored encrypted and never returned
- Keys bound to another project cannot read a source
- Disconnecting keeps the files until you delete them in Files
- An Idempotency-Key guards a retried create
Pay for what you keep
Storage
Datasets, checkpoints and exported weights are metered by the GB-hour at the monthly rate, so you only pay while you keep them.
Training tokens are metered separately when a job runs.
More in Train
Fine-tuning
Managed training on 25 open models
Training loop
DocsRL, DPO and distillation as API calls
Models
Open bases across six modalities
Read on
Connect your first source
Pick a source from the catalogue, map it, and set the refresh. The first file is ready in minutes.