I had the opportunity to participate in the AI + Digital Evidence Hackathon, held from January 15th to 17th and organized by the FĂ©nix Foundation. This incredible event brought together legal and technical experts to collaboratively tackle challenges in analyzing digitally derived evidence (DDE).

The basis for the chatbot came during the hackathon as I explored ways to streamline the application of the Leiden Guidelines. These guidelines serve as a critical framework for legal practitioners, outlining best practices for evaluating and presenting DDE in international criminal courts and tribunals. They address evidence derived from digital technologies such as videos, satellite images, call data records, and intercepted communications.

Building on the insights from the Leiden Guidelines and the incredible resources shared during the event, I created a chatbotb to tackle the growing reliance on digital evidence in legal contexts. By integrating advanced technologies like Falcon-7B Instruct LLM, Chroma, and sophisticated document processing tools, I tried to design a solution that could easily analyze and manage complex legal documents.

The concept

At its core, the chat leverages the Falcon-7B Instruct model, a state-of-the-art large language model (LLM) designed for natural language processing tasks. Falcon-7B is known for its ability to provide contextually accurate, concise, and insightful responses to complex queries. Its performance stems from its extensive training on a vast dataset, enabling it to understand the nuances of legal language and adapt to diverse scenarios.

Complementing Falcon-7B is Chroma, a powerful vector search engine integrated into the chat. This technology enables similarity-based question answering, ensuring that relevant sections of documents are retrieved quickly and accurately. By indexing legal documents into a vector database, Chroma allows matching queries with the most pertinent information, making document navigation both fast and intuitive.

The chatbot’s document processing capabilities are enhanced by robust file handling systems. Lawbotica can analyze a range of formats, including PDFs, text files, Word documents, and HTML. This flexibility ensures that users can seamlessly upload and extract insights from their legal materials without worrying about compatibility issues. These features are accessible through an intuitive Gradio app interface, allowing users to easily interact with the chatbot, upload documents, and receive insights in a user-friendly environment.

Fine Tuning the model

The lagal chatbot comes preloaded with a Q&A dataset for quick legal references. This dataset is easily expandable, allowing users to tailor it to their specific needs by adding new questions and answers. Additionally, six key legal documents, including PDFs such as the Leiden Guidelines and case summaries, were incorporated during development to fine-tune its capabilities and ensure robust legal analysis.

By using a document analysis feature I eliminated the need for manual review by extracting key insights from legal PDFs and other formats. Powered by Falcon-7B, the chat provides concise, context-aware responses to even the most challenging legal queries, while its integration with Chroma ensures fast and accurate retrieval of relevant document sections.

Limitations

While Lawbotica demonstrates significant potential in simplifying legal document analysis, it comes with a few limitations to consider:

  1. Educational Tool: Lawbotica is designed as a proof-of-concept and should not be used as a substitute for professional legal advice. It is intended to demonstrate the capabilities of AI in legal contexts but is not certified for actual legal practice.

  2. Model Accuracy and Hallucination: Like all AI language models, Falcon-7B Instruct is prone to hallucination—producing responses that seem plausible but are factually incorrect or unsupported. Users should critically assess the outputs and cross-verify critical information.

  3. Limited Scope: The system is optimized for specific types of legal documents and evidence, such as those outlined in the Leiden Guidelines. It may not perform as effectively with other forms of legal data or non-standard document formats.

  4. Resource Intensive: For optimal performance, Lawbotica benefits from systems with robust computational resources, particularly a GPU. Running the application on lower-spec machines may result in slower processing times.

  5. Data Sensitivity: Although data is processed locally, depending on the hosting setup for the Gradio app, some data might be transmitted to external servers. Users should avoid uploading sensitive or confidential legal information.

Contributing

If you’re curious about the project or have ideas for new features, we encourage you to check out our GitHub repository. Feel free to explore the code, suggest improvements, or even contribute directly.