[FEATURE] Introduce Schema Validation and Confidence Scoring Layer for LLM Extraction Reliability

[Lochit-Vinay]

📝 Description

FireForm currently relies on LLM-generated structured outputs from unstructured incident reports. In practice, these outputs are not always consistent — they can be incomplete, slightly malformed, or contain incorrect values.

This can cause issues in downstream steps like PDF auto-fill and affects the overall reliability of the pipeline.

This issue focuses on improving the reliability of the LLM → structured JSON → PDF flow.

💡 Rationale

LLM outputs are not guaranteed to strictly follow a schema. Some common issues observed:

• Missing required fields
• Incorrect data types
• Partially structured or noisy responses

Right now, there is no dedicated validation layer to catch or handle these issues before the data is used further.

Adding a validation + scoring layer would help ensure safer and more reliable processing.

🛠️ Proposed Solution

Introduce a lightweight validation and scoring step after LLM extraction:

• Schema-Based Validation

  • Use Pydantic models to enforce required fields and types
  • Flag missing or invalid values

• Confidence Scoring

  • Assign a basic confidence score per field (e.g., based on parsing reliability or fallback usage)
  • Lower confidence for fallback/cleaned/uncertain values

• Structured Error Handling

  • Standardize validation errors
  • Improve debugging and visibility into failures

This can be implemented as a modular component in the existing extraction pipeline.

✅ Acceptance Criteria

• Extracted JSON validates against a defined schema
• Missing/invalid fields are clearly flagged
• Field-level confidence scores are included
• Pipeline continues gracefully even on validation issues
• No breaking changes to current workflow

📌 Additional Context

This would be an initial implementation focused on improving robustness.
It can later be extended with more advanced validation rules or human-in-the-loop correction if needed.

This can serve as a foundation for improving extraction quality during the GSoC development phase.

[Lochit-Vinay]

I’d like to work on this and can start with an initial implementation. I’ll raise a PR shortly.

[Arijit429]

I've been working on this area — I have a few related PRs already open:

• feat: add weighted confidence scoring for extraction validation #424 — Weighted confidence scoring for extraction validation
• feat: add validation and review pipeline for extraction workflow #417 — Validation and review pipeline for extraction workflow
• Add requires_review flag to detect incomplete LLM extraction #349 — requires_review flag for incomplete LLM extraction

These cover schema-based validation, field-level confidence scoring, and flagging incomplete outputs. Happy to coordinate so we don't duplicate work.

[Lochit-Vinay]

Thanks for sharing this!
I’ll review the existing work to avoid duplication and focus on a minimal, incremental part of the validation layer for an initial implementation.

[Lochit-Vinay]

Opened a PR for an initial implementation here: #455
This focuses on a minimal schema validation layer to keep the change incremental and easy to integrate with the existing pipeline.