FormsiQ – Mortgage Transcript Field Extractor

=====================================

FormsiQ is an AI-powered application that extracts structured field data from mortgage call transcripts using Google Gemini. It supports mapping fields to the official 1003 Uniform Residential Loan Application (URLA) format and includes confidence scoring, error handling, and unmapped field logging for full traceability.

---

Features

• Structured field extraction from free-form transcripts
• Confidence scores assigned for every extracted field (0.0 to 1.0)
• Official 1003 field name mapping with fallback logging for unmatched fields
• Streamlit UI for interactive testing and demo
• FastAPI backend with simple JSON POST interface
• Handles vague or ambiguous inputs without hallucinating data
• Logs unmapped or nonstandard field names to a debug file
• Synthetic test prompts included for edge-case validation

---

How it Works

Input

• User pastes or selects a mortgage call transcript

Extraction

• Gemini 1.5 generates a JSON response using a carefully crafted prompt
• Output includes:
  • field_name
  • field_value
  • confidence_score (0.0 to 1.0)

Output

{
  "fields": [
    {
      "field_name": "Loan Amount",
      "field_value": "$350,000",
      "confidence_score": 1.0
    }
  ]
}

---

Fallback Handling

• If a Gemini-generated field name doesn’t match the 1003 form:
• • It’s stored in unmapped_fields_log.txt
• • This helps you track hallucinated or nonstandard values during testing

Sample Output Format

{
  "fields": [
    {
      "field_name": "Loan Amount",
      "field_value": "$300,000",
      "confidence_score": 1.0
    },
    {
      "field_name": "Pet Name",
      "field_value": "Luna",
      "confidence_score": 0.6
    }
  ]
}

---

Project Structure

├── app.py                   # Streamlit 
├── utils/
│   ├── api.py               # FastAPI 
│   ├── extractor.py         # Field extraction logic
│   ├── field_map.py         # 1003 form mapping dictionary
│   └── sample_prompts.py    # Test prompts
├── test_result/             # run_test file gets saved
├── run_tests.py             # Script for bulk testing edge cases
├── unmapped_fields_log.txt  # Unmapped fields are logged in unmapped_fields_log.txt for debugging
├── requirements.txt
└── README.md

---

How Confidence Works

• Pattern Matching: Detects direct keyword-value relationships (e.g., “I need a loan of $300,000” → 1.0).
• Sentence Structure: Evaluates certainty, hesitancy, or speculation in phrasing (e.g., “maybe around $300k” → 0.65).
• ≤ 0.5 → Vague, speculative, or conflicting information
• Proximity & Clarity: Scores higher when field-value pairs are close and unambiguous in context.

---

How to Run

# 1. Clone the repo and install dependencies:
- git clone https://github.com/rishabh15b/FormsiQ
- pip install -r requirements.txt

# 2. Add your Gemini API key to a .env file:
- GEMINI_API_KEY=your-api-key-here

# 3. Run the backend:
- uvicorn utils.api:app --reload

# 4. In another terminal, run the Streamlit app:
- streamlit run app.py

---

Important Notes

• Only clearly stated fields are extracted — no guessing.
• Confidence score is dynamic (0.0–1.0) based on sentence clarity.
• Hallucinated or irrelevant fields are filtered and logged.
• Mapping is strictly based on 1003 form + fallback for custom fields.
• Output is clean, structured, and directly usable.
• Edge and vague cases are synthetically tested.

---

Author

• Developed by Rishabh Balaiwar. Feel free to reach out for questions or collaboration opportunities!

---

Contact

• For support or inquiries, contact me at:
• Email: rbalaiwar@gmail.com
• GitHub: Your GitHub Profile