Docs·More

Compare

XERJ Reader next to general LLM APIs, typed-answer APIs and field-extraction APIs. Facts about other products are from their public pages, as of November 2026.

Three kinds of API can pull an answer out of a document. They differ in what they return, so they fit different jobs. Facts about other products come from the sources at the end of this page, as of November 2026. A dash means the sources do not say.

At a glance #

At $0.01, XERJ Reader is about 4x below Jev's $0.042 and 10x below Clairic's $0.10 preview price.

Against LLM extraction #

A general LLM generates text token by token. Asked for an invoice total, it writes one. The value can be paraphrased (48200, €48.2k) or invented, and it comes without offsets, so checking it means searching the document yourself. You also pay for the output tokens.

XERJ Reader does not write. It returns 48,200 EUR and the characters 533–543 it came from, or none. There is nothing to paraphrase and nothing to invent.

Use an LLM when the answer is not written in the document: a summary, a draft, an explanation, or a question that needs world knowledge.

Against typed-answer and choice APIs #

A typed-answer API such as Jev answers yes/no, picks one of up to 255 options you list, or gives a score, about a state of up to 32k tokens.

That fits decisions among labels you define in advance: routing a ticket, classifying a message, scoring a lead. It does not fit values you cannot list in advance. A date, an amount, a name or an ID is not one of your options until you know it.

XERJ Reader treats every substring of the document as a possible answer. Dates, amounts, names, IDs and clauses work without listing them, and each comes with its position.

Use a typed-answer API when the answer is a label from your own list, or a score.

Against field-extraction APIs #

A field-extraction API such as Cloudflare Clairic takes a state and a field map and returns a value per field. Its field types are text, number, date, time, datetime, duration, list, estimate and forecast. Text answers carry character spans. Each field has a confidence, a null probability and a null threshold you can set. In the published examples, values that are worked out from the text, such as a sum, a date or a duration, come without a span.

XERJ Reader's /v1/extract also reads many fields in one call. Every answered field carries its evidence span. Values come back exactly as written, so 48,200.00 EUR stays a string until your code parses it.

Use a field-extraction API when you need values computed or normalised by the model, estimates and forecasts with intervals, or image input.

Where XERJ Reader is not the right tool #

Being clear about this saves you a migration:

• No images. XERJ Reader reads text. Scanned documents need OCR first.
• No computed values. It reads 48,000.00 EUR and 200.00 EUR; it does not add them. Do arithmetic in your code. XERJ Extract, coming soon, turns fields into checked programs with evidence.
• No normalisation. Dates and amounts come back as written in the document.
• No labels from outside the document. For classification into your own categories, use a typed-answer API or a classifier.
• No estimates or forecasts.

Sources #

Facts about other products, as of November 2026:

• Jev (TypeSafe): typed answers (yes/no, choice up to 255 options, score) about a state of up to 32k tokens; $0.042 per 1M input tokens, output free.
  • docs.typesafe.ai/introduction
  • docs.typesafe.ai/api
  • pydantic.dev/docs/ai/models/typesafe
• Cloudflare Clairic: field extraction from a state plus a field map; field types, spans, confidence and null threshold as described above; preview price $0.10 per 1M input tokens, output free.
  • blog.cloudflare.com/clairic-extraction-model
• General LLM APIs: they generate text token by token and bill output tokens. This describes the category, not one product.

Prices and features of other products change. Check their pages before you decide.