[FEAT]: Adding line snippets from transcript where each field is found as an Audit Trail

[Cubix33]

📝 Description

When the AI fills out a form, it currently just drops the answer into the PDF without explaining where it got the information. We need a way to see the AI's "homework." This feature makes the AI extract the exact quote from the audio transcript that it used to find the answer. It then creates a separate Audit Trail text file alongside the filled PDF so humans can easily verify the AI's work.

💡 Rationale

FireForm handles high-stakes emergency reporting. If an incident report is ever questioned in court or by a fire chief, officials must be able to see the raw text that justified the AI's decision. We cannot have a "black box" where data magically appears.

🛠️ Proposed Solution

We will force the LLM to output its answers as a strict dictionary containing both the final answer and the source quote. Then, we will generate a clean .txt file logging this data.

• Logic change in src/llm.py: Update build_prompt to demand a strict JSON format with "value" and "quote" keys.
• Logic change in src/llm.py: Add a safety check in add_response_to_json to handle cases where the LLM mistakenly returns a Python list instead of a string (joining lists with a ; to prevent crashes).
• Logic change in src/filler.py: Extract only the "value" to fill the PDF fields, then write a formatted _audit.txt file that lists every field alongside its AI value and transcript source quote.

✅ Acceptance Criteria

• Prompt Adherence: The LLM successfully returns both the value and the quote for every field.
• Crash Prevention: If the LLM hallucinates a JSON array (e.g., ["John", "Jane"]), the system catches it and converts it safely without throwing an AttributeError.
• Audit File Generation: A clean _audit.txt file is saved in the same directory as the output PDF, containing readable timestamps, fields, values, and quotes.

📌 Additional Context

We specifically chose to generate a .txt file instead of drawing a new page inside the PDF. This avoids adding heavy new dependencies (like reportlab) to our Docker image, keeping the application lightweight, fast, and completely free of external library bloat. Furthermore, a .txt file is much easier for a fire department's database system to parse later compared to extracting text from a PDF graphic layer.

[Cubix33]

Contains changes made in the pull requests for bugs #173 and #135 for smoother execution.