[ENHANCEMENT] : Optimize LLM.main_loop() for batch extraction of fields

[dibyajyoti-mandal]

Description

Currently, the LLM extraction pipeline in src/llm.py fires a separate HTTP request
to Ollama for every field in the form via LLM.main_loop(). A form with 20 fields
produces 20 sequential round-trips to the local model, making the pipeline slow,
fragile, and wasteful. There is also no structured output contract the LLM returns
raw strings, and the -1 sentinel for missing values can silently end up written
into a PDF field.

Proposed Solution

Replace the per-field loop with a single batch extraction call LLM.extract_all() that sends all field names to Mistral at once and receives a single JSON object
containing every field value in one shot.

Suggested Implementation

New method: LLM.extract_all()

def extract_all(self, fields: list[str], transcript: str) -> dict:
    prompt = f"""
    SYSTEM PROMPT:
    You are an AI assistant designed to extract information from transcribed voice
    recordings and return the results as a structured JSON object.

    You will receive a list of field names and a transcript. For each field, identify
    its value in the transcript and include it in the JSON response.

    Rules:
    - Return a single JSON object where every key is a field name from the list.
    - If a field is plural and multiple values are found, return them separated by ";".
    - If a value cannot be found in the transcript, return null for that field.
    - Return JSON only. No explanation, no markdown, no extra text.

    ---
    Fields to extract: {json.dumps(fields)}
    Transcript: {transcript}
    """
    response = self.client.chat(
        model="mistral",
        messages=[{"role": "user", "content": prompt}],
        format="json"
    )
    return json.loads(response.message.content)

Fallback for backward compatibility

The old per-field loop is retained as LLM._legacy_per_field_extract() and called
automatically if the batch method fails (e.g. on older Ollama versions that do not
support JSON mode):

Benefits

• Performance — eliminates N sequential Ollama round-trips, replacing them with
  a single call regardless of form size
• Safety — null for missing values is unambiguous and handled cleanly by the
  downstream Validator, preventing sentinel strings from reaching the PDF
• Backward compatibility — legacy per-field loop is preserved as a fallback,
  so deployments on older Ollama versions are unaffected

Files Affected

• src/llm.py — primary change, new extract_all() and _batch_extract() methods
• tests/test_llm.py — new unit tests for batch extraction, JSON parsing, plural
  handling, null sentinel, and fallback behaviour