Skip to main content

Customizing AI document extraction & prompts WIP

LITEFile includes an automated AI extraction engine designed to inspect uploaded court PDF documents and recommend the appropriate Court, Case Category, Case Type, and Docket Number.

This guide explains how court partners and developers can customize extraction hints, field definitions, model tiers, and private LLM gateways.

1. How the extraction pipeline works​

The extraction logic lives in:

  • efile_app/efile/utils/llms.py: Core OpenAI API wrapper, model selection, and token management.
  • efile_app/efile/views/session_api.py: Jurisdiction-specific deduction hints (llm_hints) and field schemas (llm_fields).
  • efile_app/efile/services/document_uploads.py: Integration with file uploads and payload guessing.

2. Customizing jurisdiction deduction hints (llm_hints)​

In efile_app/efile/views/session_api.py, the llm_hints dictionary supplies court case classification rules to guide the LLM's reasoning:

# efile_app/efile/views/session_api.py

llm_hints = {
    "illinois": """
        However you should always attempt to deduce "case_category" and "case_type", using the following information:

        * Chancery (CH): Specific Performance, Injunction, Mechanics Lien Foreclosure
        * Criminal Felony (CF) or Criminal: Petition to Expunge or Seal
        * Dissolution with Children (DC) or without Children (DN) (NOTE: Dissolution means Divorce)
        * Misdemeanor (CM)
        * Eviction (EV) NOTE: Eviction may also be called Forcible Entry and Detainer
        * Family (FA): Petition for Parentage, Visitation, or Custody
        * Guardianship (GR): Guardianship of Minor or Person with Disability
        * Law Magistrate (LM): Claims for money over $10,000 up to $50,000
        * Miscellaneous Remedy (MR): Change of Name, Administrative Review
        * Order of Protection (OP): Order of Protection, Stalking No Contact, Civil No Contact
        * Probate (PR): Administration of Decedent’s Estate
        * Small Claims (SC): Claims for money $10,000 or less
    """,
    "massachusetts": """You should always attempt to deduce "case_category" and "case_type".""",
    "vermont": """You should always attempt to deduce "case_category" and "case_type".""",
    "default": """You should always attempt to deduce "case_category" and "case_type".""",
}

Adding hints for a new state:​

To add or refine hints for your state, add an entry keyed by your jurisdiction identifier (e.g., "massachusetts" or "california"). Include statutory abbreviations, case code prefixes, and common legal synonyms.

3. Customizing target fields (llm_fields)​

The llm_fields dictionary defines the target JSON schema and field descriptions passed to the model:

# efile_app/efile/views/session_api.py

llm_fields: dict[str, dict[str, str]] = {
    "illinois": {
        "court name": "The name of the court that this form is filed in, often is the county of the court.",
        "filing type": "The formal title of the filing being made",
        "case category": "The high level category of this case",
        "case type": "The type of legal case this form is a part of",
        "docker number": "The unique identifier for this case in court. Also referred to as the case number",
    },
    "default": {
        "court name": "The name of the court that this form is filed in.",
        "filing type": "The formal title of the filing being made",
        "case category": "The high level category of this case",
        "case type": "The type of legal case this form is a part of",
        "docker number": "The unique identifier for this case in court. Also referred to as the case number",
    },
}

4. LLM provider configuration & privacy​

LITEFile uses standard OpenAI-compatible endpoints, allowing you to use:

  • Commercial cloud models: OpenAI (gpt-4o-mini, gpt-4.1-mini, gpt-5-mini), Anthropic Claude, Google Gemini.
  • Self-hosted / local LLM gateways: vLLM, Ollama, LiteLLM proxy, or Azure OpenAI for strict court data privacy compliance.

Environment variables:​

# OpenAI or Private Gateway
OPENAI_API_KEY="sk-..."
OPENAI_BASE_URL="https://api.openai.com/v1/"  # Or http://localhost:8000/v1/

Model selection & tiers:​

In llms.py, models are arranged in three performance tiers:

  • Small (default): gpt-4o-mini, gpt-4.1-nano, gemini-2.5-flash-lite, claude-3-5-haiku
  • Medium: gpt-4o, gpt-4.1-mini, claude-3-7-sonnet
  • Large: gpt-4o, o3, claude-3-7-sonnet

The system automatically verifies model availability against the /models endpoint of your configured provider and falls back gracefully.