An agent for every kind of document you receive
Textikraft turns two or three examples and a short conversation into a versioned, tested, billable extraction API.
Train it by talking to it
Drop in a few samples and say what matters, the way you would brief a new colleague: "These are supplier invoices. Take the invoice number, date, vendor and line items, and check the total."
The agent reads your samples, proposes the output fields and shows what it extracted from each one. Not right? Say so and it tries again. When an answer is correct, approve it. Approved answers become the test set that decides whether the agent is good enough to publish.
- Choose your own output format, or let it propose one
- Side-by-side view of the original and the result
- Questions back to you when something is ambiguous, such as day and month order in a date
Everything is explicit
You see the fields, the instructions the agent learned, the rules it wrote and the score on every sample. Nothing about how it reads your documents is hidden from you.
The AI writes the rules, then steps back
For text-based documents, the agent turns what it learned into plain extraction rules and arithmetic checks. In production those rules run first, with no AI call and no AI cost. The model is only asked about the fields the rules cannot read, and every result still goes through the same checks.
Rules first
Labels, positions and patterns that held in every sample become rules that run in milliseconds.
AI for the gaps
Only fields without a rule go to the model. The model never overwrites a value a rule already verified.
Checks on everything
Schema check, a check that each value really appears in the document, and arithmetic checks such as line items adding up to the total.
A second look when weak
If a result fails checks, a stronger model gets one more try before anything reaches you.
Only good agents go live
Every build is tested against your approved answers and scored field by field. You set the accuracy bar for publishing; the platform enforces a minimum. Each version keeps the exact samples, answers, format and instructions it was built from.
Version history
See accuracy for every version on a chart. If a retrain scores lower than the live version you are warned, and the live version keeps serving your API.
Roll back or retrain from any version
Make an older version live again in one click, or restore its training set and keep improving from there.
Production, bulk, and everything around it
One endpoint, many layouts
Group the different looks of a document type under one agent. Each document is matched to a layout automatically. If none matches, you are told, and charged only the minimum.
Batches and ZIPs
Send hundreds of documents at once and download the results as JSON or CSV.
Review and correct
People confirm or fix results in a side-by-side screen. Corrections give you true accuracy and can join the agent's test set.
Webhooks and idempotency
Signed webhooks tell your system when a document is done. Idempotency keys make retries safe.
Sessions and usage
For every agent: documents uploaded, credits spent, overall accuracy, accuracy for each document, and a detailed log.
Teams and access
Roles, per-person menu access, two-step sign-in and an audit log. Your account manager can see your usage but not your documents' contents.