Dashier’s AI Ingest turns a plain-language description into an editable data-source draft, then validates proposed rows before you choose whether to import them.
Dashier organizes data as organization → project → source → record. A source is a structured collection of records, with fields that define what each record can contain. When the shape of your data is clear in your head but not yet configured in Dashier, AI Ingest can help translate a description into a first draft. When the data already exists in a clean spreadsheet or JSON file, the deterministic CSV or JSON importer is usually the more direct choice.
AI produces a proposal; it does not silently create sources or write records. You inspect, edit, validate, and explicitly run an import using Dashier’s normal scoped application routes.
What AI schema generation does—and does not do
AI Ingest is a prompt-based assistant for data-source design and bulk import preparation, not a machine-learning model trainer. You describe what you are organizing and, if needed, provide example rows. The assistant can draft source names, field names and types, required and unique settings, and proposed records. You can revise that draft in the interface before committing anything.
For instance, a team setting up a small contacts directory might ask for a Contacts source with an email address, a name, and an optional organization. Review whether email should be required or unique, whether the chosen types fit your use, and whether supplied records are accurate. A generated draft is not a guarantee of a correct schema or clean data.
It is not a replacement for source permissions, the ordinary field builder, or a review of imported values. Learn about the workflow on the AI Ingest feature page, and review how sources work in Data Sources documentation.
Connect an OpenAI-compatible provider
Before opening AI Ingest, configure an AI provider in your organization’s AI settings. You can use your own reachable public HTTPS OpenAI-compatible chat-completions provider. Enter the provider’s API key and model name. The provider must support tool calling for this workflow; compatibility with every provider or model that accepts chat-completions requests is not guaranteed.
Saving provider settings and testing the connection are separate actions. Save the configuration, then run the test explicitly. A successful save does not establish that the provider is reachable, the key is valid, or the model supports tool calling. If the test fails, check the endpoint, key, model identifier, and tool-calling support.
AI Ingest is opened from Data Sources in the relevant project. The organization and project context matter: sources belong to a project, and your access to that project controls which actions are available. See Getting Started for the organization and project setup, and Permissions for access rules.
Write a prompt that gives a useful draft
A prompt can be up to 20,000 characters. Be specific about the purpose of the data, the source boundaries, and what should happen to missing values. Avoid asking for a large, underspecified system when a narrow first draft will be easier to review.
A useful schema-only prompt could be:
Create one source called Contacts for a small team directory. Include
full_nameas required text,organizationas optional text. Draft the schema only. Do not generate records. Do not invent missing values.
This tells the assistant what the source represents, identifies the intended types and constraints, and makes clear that the request is for a schema rather than sample data. The output is still a draft: review each name, field type, required flag, and uniqueness choice in Dashier. For an existing source, describe the columns in the supplied batch and how they correspond to existing fields. Don’t assume that wording alone will resolve ambiguous values or make an incompatible row valid.
If a field’s valid choices are known, state those choices clearly. Enum values can be configured inline, or you can select an existing reusable static list. A reusable list is helpful when multiple sources need the same controlled values; inline choices may be appropriate for a source-specific set. Choose and verify the actual values rather than relying on the assistant to infer a canonical list.
Example: schema first, records second
For the contacts example, keep schema design separate from the data you intend to import. First request the schema only with the prompt above. Review and edit the draft, click Re-validate, and explicitly run it to create the empty Contacts source. Then open AI Ingest again for a separate import request. For example, your supplied rows might be:
full_name,email,organization
Ada Example,ada@example.com,Northstar Studio
Lin Example,lin@example.com,
Name the destination explicitly: “Prepare these two supplied rows for import into the existing Contacts source. Keep the blank organization empty. Do not create another source or invent additional rows.” Include the rows with that request, then review and validate the returned import plan before running it.
These are illustrative rows for this workflow, not real customer data. The blank organization cell for Lin is intentionally blank; do not ask the assistant to fill it by guessing. A cautious instruction is: “Use only the rows I supplied. Do not generate records or invent missing values.” Such instructions help express the intended handling, but they do not guarantee that a model will always follow them perfectly. Inspect the proposed records, especially blanks, normalized values, and field matches, before import.
Choose the right workflow and size
From a project’s Data Sources area, AI Ingest can draft a new source, draft several sources, or prepare records for an existing source. The supported sizes differ by workflow:
- One new source: draft 1–100 fields and 0–500 initial rows.
- Multiple new sources: draft 1–25 sources, with 0–500 initial rows for each source.
- Existing source: prepare an import of 1–2,000 rows.
- Matching existing records: prepare 1–500 updates, matched using a unique field or a generated ID only.
Use a new source when the data represents a distinct collection and its fields are not configured yet. Use multiple sources only when the information really belongs in separate collections—for example, a simple inventory source and a separate supplier source—not merely because a prompt is long. For a known source and a larger batch, split the work into manageable, reviewable imports within the supported limit.
An AI draft does not create reference fields. If your design needs a relationship between sources, create or edit the source fields afterward with Dashier’s normal field builder, then verify the relationship and records. For example, a supplier relationship on an inventory source is a deliberate schema decision; it should not be inferred from a text column that happens to contain a supplier name.
Existing-record matching has a narrower contract than ordinary field mapping. Updates must use a unique field or a generated ID; a name that merely looks distinctive is not a safe substitute. Check which field will identify each target record and review the proposed matches before running updates. If the identifier is missing, duplicated, or uncertain, resolve that first rather than hoping the import can identify the intended record.
Review, edit, re-validate, then run
Treat the returned draft as an editable proposal. Inspect source names, fields and their types, required and unique settings, enum values or list choices, and every proposed row. Remove anything you do not intend to import, correct mappings, and supply missing information only from a trustworthy source. For updates, carefully verify the match key and target records.
After the draft arrives—or after you make edits—click Re-validate explicitly. The current preview may not automatically pick up newly received draft state. Re-validation is the explicit step to refresh checks against the current draft; do not rely on an older preview appearing current. If validation reports issues, edit the draft and re-validate again. Only use the run action when the draft is valid and the preview represents the version you reviewed.
This sequence matters because drafting and importing are different operations. The AI proposes a structure and values; validation checks the current proposal; your explicit run action requests the write. Normal validated, project-scoped routes apply the changes. The feature is not an arbitrary autonomous agent with permission to explore your backend or make unrequested changes.
For a clean file with settled columns, use the CSV and JSON import guide instead. A conventional importer is repeatable and makes the input-to-field mapping direct; AI is most useful when you want help shaping an initial schema or preparing data that needs interpretation. See the API reference for the broader programmatic interface.
Privacy and operational limits
The source or list context and your prompt are sent to the configured provider so it can prepare the draft. Do not include secrets, credentials, access tokens, or sensitive personal information in prompts. Minimize personal data, and provide only the context needed to produce a useful proposal. Use a provider you are authorized to send that information to, and follow your organization’s data-handling rules.
Current limits: you must manually click Re-validate to avoid relying on a preview that may miss newly received draft state. There is no MCP integration, no arbitrary automatic backend agent, and no arbitrary JSON Schema generation. Tool-calling requirements mean compatibility with any particular model or provider is not guaranteed. Multi-source and record writes can partially succeed rather than behaving as one guaranteed atomic import. Check the results for each source or batch before retrying; an unexamined retry can duplicate work or apply changes twice. If a run has mixed outcomes, identify what succeeded and what remains before preparing a retry.
Frequently asked questions
Does AI Ingest immediately create sources or records?
No. It generates an editable draft. You review it, explicitly re-validate it, and choose to run the import only when the current draft is valid.
Can it generate relationships between sources?
Not as reference fields in the AI draft. Add relationships later through the ordinary field builder and validate the relevant data and references there.
Should I use AI for a clean CSV file?
Usually not. If the columns and values are already structured and the target fields are known, use the deterministic CSV or JSON importer. Choose AI Ingest when you need a schema or mapping proposal that you will inspect and edit.
What if an import only partly succeeds?
Review the per-source or per-record results and check the project before retrying. Writes are not guaranteed to be atomic across a multi-source or record import, so determine exactly what succeeded and avoid blindly resubmitting the entire batch.
A practical checklist
Before starting, confirm the organization and project, and that provider settings are saved and tested. Keep the prompt focused, respect limits, and omit unnecessary sensitive information. Review fields, constraints, values, and update identifiers. Re-validate when the draft arrives and after edits; run only when valid. Inspect results before preparing any follow-up batch.
The best outcome is not a schema accepted without thought. It is a clear, maintainable source definition and a reviewed batch whose values belong in that source. Dashier’s AI can speed up the first draft, while the normal field builder, validation steps, permissions, and explicit import action keep the schema and data under your control.
