A prompt is the instruction given to the language model: what the document is, which fields to return and in which format. ChronoScan keeps the prompts of the AI configurations of each job and offers two tools to work on them.
AI Prompts window (Indexing tab, Batch operations)
Click AI Prompts with a document selected to try prompts on it. Pick one of the example prompts or write your own, choose the provider and the model, and run it: the request is sent with the text (and optionally the image) of the current document and the answer is shown in the window and in the ChatGPT output panel (Panels menu). Nothing is written to the fields: it is a laboratory to tune the wording before copying it to an AI configuration.
Prompt library
The system prompt and the JSON template of each AI configuration can be saved in the prompt library of the job and reused in other configurations or providers. The library is kept in the job folder (openai_prompts_cfg.json).
Writing a good extraction prompt
Give the business context in one sentence (supplier invoices of a Spanish company, delivery notes...).
Define each field, its format (dates dd/MM/yyyy, decimal separator) and what to return when it is missing.
Describe the line items as a table with the column names of the capture grid.
Do not include examples with real identifiers: the model copies them when the document does not contain them.
Prefer the structured response mode (schema derived from the JSON template) to avoid truncated or invented JSON.
Prompts can also be called from scripts for special cases (summaries, business checks) with ChronoDocument.GPTRequest and GeminiRequest; see ChronoDocument.
