Format Workflows

How to fill each of the four form formats. Read the section that matches the file in hand.

The PDF Tools connector (labeled "PDF Tools - Fill, Sign, Merge, Split, Extract") is the workhorse for PDFs. Its tools require absolute paths on the user's local machine, not container paths — pass the real file path the user gave you.

A. Fillable PDF (AcroForm)

Best case — the PDF has named form fields you can set directly.

Read the fields: call read_pdf_fields on the PDF. This returns every field name, type, and current value. This tells you the exact field names to target (they're rarely pretty — e.g., topmostSubform[0].Page1[0].LegalName[0]).
Ensure the mpm profile exists: call list_profiles. If there's no mpm profile, seed one: read assets/mpm-profile.json, build a flat {field_name: value} map from the non-null values, and call save_profile with profile_name: "mpm". (Saving once makes every future fill a one-liner.)
Map profile → form fields: the saved profile keys won't match the PDF's field names. Build the mapping by matching each PDF field name/label against the aliases in mpm-profile.json. Values you can't map, or whose profile value is null, stay out.
Fill: call fill_with_profile with pdf_path, output_path (outputs folder), profile_name: "mpm", and an additional_data object carrying the specific field-name→value pairs you mapped in step 3 (this overrides/supplements the profile). If it's simpler, skip the saved profile and pass everything through additional_data.
Leave human/blank fields alone: do not fill signature, date, or certification-selection fields.
Verify: re-read fields or render the page to confirm values landed in the right boxes.

B. Flat / scanned PDF (no form fields)

The PDF looks like a form but read_pdf_fields returns nothing fillable. You overlay text.

Understand the layout: call get_page_analysis (and/or render_pdf_page) to see where labels and blank lines sit. search_pdf_text can locate a label's coordinates so you can place the value just after/below it.
Overlay text: use the connector's apply_text to place each value at the right (x, y) on the right page. For many fields, the bundled scripts/fill_flat_pdf.py is faster — it takes the PDF plus a JSON list of {page, x, y, text} placements and writes a filled copy in one pass.
Coordinate system: PDF origin is bottom-left, y increases upward, points (72 = 1 inch). If get_page_analysis gives top-left coordinates, convert: y_pdf = page_height - y_top. Always render the result and eyeball it — flat-PDF placement is the most error-prone format.
Signature/cert fields: leave blank.
Verify: render each filled page and confirm text sits on the blanks, not over labels.

If placement proves too fiddly or the scan is low quality, fall back: reproduce the form as a clean .docx (Format C) and note to the user that you rebuilt it rather than annotating the scan.

C. Word (.docx) form

Use the docx skill (read its SKILL.md for the mechanics of editing Word files).

Open the document and inspect its structure — agency Word forms are usually tables where the left cell is a label and the right cell is the answer, or inline "Name: ____" runs.
For each labeled cell/blank, match the label to a profile field (via aliases) and insert the value into the answer cell/run, preserving the surrounding formatting.
Leave signature lines, date lines, and certification checkboxes for the human.
Save the filled .docx to outputs. Verify by re-reading the document text.

D. Table embedded in the RFP body

A certification block or info table living inside the solicitation document itself (not a separate form).

Locate and extract the exact table/block (with the docx or pdf skill depending on the source).
Decide the return format with the user's intent in mind: usually cleanest to reproduce the table in a standalone .docx labeled with the solicitation/agency, fill MPM's values, and hand that back. If the user wants the whole document returned with the block filled in place, do that instead.
Fill via alias matching; leave human-affirm items marked.
Save to outputs and verify.

After any format

Keep the three running lists the SKILL.md asks for: (1) fields filled, (2) fields left blank for lack of data, (3) certification/signature items for the human. Report them back — the "left blank" list is the user's actionable to-do.

4.5 KiB Raw Blame History