Pydantic refactor

[Cubix33]

Closes #148

🚀 Description

This PR fundamentally overhauls the core extraction logic in src/llm.py by integrating Pydantic and Ollama's native JSON Structured Outputs.

Previously, main_loop iterated sequentially through every field, making N API calls to the LLM and relying on fragile string parsing (replace('"', "")) to format the output. This was slow, prone to hallucinations, and broke easily if the LLM wrapped its answer in unexpected formatting.

This update introduces a single-pass, strictly typed architecture:

1. Dynamically generates a Pydantic Model (create_model) based on the target PDF fields.
2. Converts the Pydantic model into a JSON Schema.
3. Passes the schema directly to Ollama via the format API payload parameter.

This reduces extraction time by drastically cutting API calls and guarantees 100% reliable data structures.

🛠️ Changes Made

• Added pydantic>=2.0.0 to requirements.txt.
• Refactored src/llm.py:
  • Removed outdated add_response_to_json and handle_plural_values string parsers.
  • Added build_schema() to map complex PDF field names to valid Python identifiers.
  • Added map_schema_to_json() to reconstruct the raw target fields for filler.py.
  • Updated main_loop() to execute a single, schema-enforced API call.

🧪 How to Test

1. Run pip install -r requirements.txt to install Pydantic.
2. Ensure Docker and Ollama are running.
3. Execute make exec.
4. The console will report [LOG] Extracting all fields using Pydantic structured output... and complete the entire PDF mapping in a single, fast operation.

✅ Checklist

• Tested locally within Docker.
• Verified JSON outputs map perfectly to filler.py expectations.
• Maintains compatibility with the Stateful Resumption logic.

[Cubix33]

Closing for now to reduce load on maintainers. Will reopen after further discussion or during GSOC.