Can AI fill forms from documents? Yes, in two separate steps: extract the values, then place them.
Short answer. Yes, in two steps that should stay separate. First a model extracts the values you need from the source documents into a fixed structure; Ollama and vLLM can both constrain a model’s output to a JSON schema, so the fields arrive in a predictable format. Then a script, not the model, writes those values into the form. For PDF forms with fillable fields, libraries such as pypdf read the field names and fill them in. Scanned or flat forms, forms built on XFA, and forms whose fields have generic names are harder. A person should check every value against its source before the form is signed or filed.
01
How hard is your form?
Form type
Route
Difficulty
Fillable PDF (AcroForm)
Read the field names with pypdf, map each to a value once, then fill and optionally flatten the form.
Easy to moderate. Field names are often generic, so the mapping is hand-made once per form.
PDF built on XFA
pypdf notes that an /XFA form, a specific XML description of the form, overrides the page content, so it needs special handling.
Hard. Consider re-keying into a standard form.
Flat or scanned PDF, or paper
OCR the form, locate each blank by position, then draw the text at those positions.
Moderate to hard. Positions must be exact.
02
How extraction works
Ollama’s structured outputs, announced in December 2024, let you constrain a model’s output to a format defined by a JSON schema. vLLM supports structured outputs with the xgrammar or guidance backends. Define each field with a name, a type and, where you can, a list of allowed values, and include an explicit unknown option. The model then cannot reply in free text, and a missing value can be flagged rather than guessed.
A schema guarantees the shape of the answer, not its truth. A well-formed date can still be the wrong date. That is why the values must be checked against the source, and why each should carry a pointer to the passage or page it was read from.
03
What a person must still check
Every value against its source: names as they appear on the identity document, dates in the form’s own format, amounts and units.
Fields the documents do not support. They should stay blank and be flagged, not filled with a plausible guess.
Check boxes and choices, where a wrong tick is easy to miss.
Declarations and signatures, which only a person can give.
Sensitive identifiers such as Aadhaar and PAN numbers. Fill only the fields the form needs and keep the working copies protected.
The finished form, opened as the recipient will see it, not only in your own viewer.
04
How an office would set this up
Start with one high-volume form that has real fillable fields.
List its fields and write the schema, with the source document each value should come from.
Extract, then run automatic checks: formats, required fields, totals.
Fill the PDF with a script and keep the filled copy and the extraction record together.
Have a person review each form beside its sources. Track the corrections for a month.
Add a second form only when the first has a low and stable correction rate.
05
What ANVE does today
ANVE does not list form filling as a function. It lists question answering over the office’s own documents with the source passage shown, vision models that read scanned documents with OCR, and a Bills function that reads GST bills. A form-filling agent would be separate work. The agents ANVE builds with a paid pilot are legal research, compliance check, tender response, action follow-up, policy helpdesk and a voice front desk; they are not shipped products.
06
Questions people ask
Can AI sign the form?
No. A person must make declarations and sign. The AI can prepare values for the person to check.
What if the information is not in the documents?
The field should stay empty and be flagged for a person. Give the model an explicit unknown option so it is never pushed into guessing.
Does a JSON schema make the extracted values correct?
No. It makes them well-formed. A correctly formatted date or amount can still be wrong, so each value needs checking against its source.
Can it fill a form in Hindi?
It can if the model reads the source well and the form’s fonts support the script. Test with a real form and check how the filled text renders in the recipient’s viewer.
07
What is not real yet
No sealed ANVE unit exists; every image of the appliance on this site is a labelled render. The software runs today on a development machine. No customers yet, no certifications, no delivery date, and the final price is not announced.
Reservations are ₹999 in India or US$12 elsewhere, refundable in full until you confirm an order in writing (terms). Offices that cannot use cloud AI can apply to be one of ten design partners.
