[FEAT]: Department Profile System for Pre-Mapped PDF Templates

[utkarshqz]

📝 Description

FireForm currently extracts PDF field names as machine-generated identifiers (e.g. textbox_0_0, textbox_0_1). When these are sent to Mistral for extraction, the model has no semantic context and either returns null for all fields or hallucinates a single value repeated across unrelated fields (see related Bug #173).

A Department Profile system would ship pre-built mappings between human-readable field labels and the internal PDF field identifiers for common agency forms used by Fire Departments, Police, and EMS.

💡 Rationale

FireForm's mission is to serve real first responders out of the box. Currently:

• A firefighter uploads a Cal Fire incident form
• Mistral receives {"textbox_0_0": "", "textbox_0_1": ""}
• It has no idea what these fields mean → returns null or wrong values
• The filled PDF is blank or incorrect

With department profiles:

• The profile provides {"Officer Name": "textbox_0_0", "Incident Location": "textbox_0_1"}
• Mistral receives human-readable labels → extracts correctly
• The filled PDF is accurate

This solves the root cause of Issue #173 without requiring changes to the LLM pipeline.

🛠️ Proposed Solution

• Create src/profiles/ directory with JSON profile files
• Each profile maps human-readable field labels → internal PDF field IDs
• Add profile selector to the frontend UI (dropdown by department type)
• Pass field label mapping to LLM prompt during extraction

Profile schema:

{
  "department": "Fire Department",
  "description": "Standard Cal Fire incident report",
  "fields": {
    "Officer Name": "textbox_0_0",
    "Badge Number": "textbox_0_1",
    "Incident Location": "textbox_0_2",
    "Incident Date": "textbox_0_3",
    "Number of Victims": "textbox_0_4"
  },
  "example_transcript": "Officer Smith, badge 4421, responding to structure fire at 742 Evergreen Terrace on March 8th. Two victims on scene."
}

Profiles to implement:

• fire_department.json — Cal Fire incident report
• police_report.json — Standard police incident form
• ems_medical.json — EMS patient care report
• Logic change in src/llm.py to use profile labels in prompt
• Frontend dropdown to select department profile

✅ Acceptance Criteria

• At least 3 department profiles ship with the repo
• Profile labels are injected into the Mistral prompt
• Extraction accuracy improves for pre-mapped forms (no null output)
• Feature works in Docker container
• Documentation updated in docs/
• JSON output validates against the schema

📌 Additional Context

Related bugs this directly addresses: #173 (PDF filler hallucinates repeating values)

Related features this complements: #111 (Field Mapping Wizard — for custom PDFs not covered by profiles)

This is especially important for FireForm's stated mission as a UN Digital Public Good — the system should work correctly for real first responders without requiring technical setup.

[srijitac1]

please do check my PR

[utkarshqz]

Hi @srijitac1 , thank you for taking the initiative to implement the Department Profile system I proposed! It's great to see this moving forward so quickly.

I’ve reviewed the PR and it aligns well with the initial JSON structure I outlined. As I’ve been building out the broader Dynamic Extraction Pipeline (see PR #241 and #246), my focus has actually moved toward making these profiles even more powerful. My goal is to ensure the system isn't limited to a fixed set of fields but can instead handle any PDF length dynamically by merging these profile schemas with real-time field discovery.

I’d be happy to support your PR and see how we can integrate your static profile data into the dynamic pipeline I'm finishing up. It would make for a very robust "Zero-Configuration" system for agencies!

[srijitac1]

Hi @utkarshqz, thanks for the detailed review and insights!

I’ll go through PR #241 and #246 to better understand the dynamic extraction pipeline and how everything fits together. Then I’ll look into how my current implementation can align and integrate with that direction.

Looking forward to contributing further here! :)