DocuLegal — Legal Document Intelligence

A retrieval-augmented generation (RAG) application for asking plain-English questions about legal PDFs. It retrieves relevant passages and generates a structured answer with source references.

Features

• Upload one or multiple PDFs.
• Ask questions about clauses and compare information across documents.
• Retrieve short passages while returning larger surrounding context.
• Combine keyword and semantic retrieval with reranking.
• Display page/section references, follow-up answers, and confidence labels.

Retrieval pipeline

1. Load PDFs with PyPDFLoader.
2. Create parent chunks (1,500 characters) and child chunks (300 characters).
3. Embed child chunks using sentence-transformers/all-MiniLM-L6-v2 and store them in Chroma.
4. Combine BM25 and vector retrieval using reciprocal rank fusion.
5. Rerank candidates with cross-encoder/ms-marco-MiniLM-L-6-v2.
6. Return parent context and generate a structured response using the Groq API.

Component	Technology
Application	Python, Streamlit
Framework	LangChain
Storage	Chroma and an in-memory parent-document store
Generation	Groq API; documented model: openai/gpt-oss-120b
Response structure	Pydantic schemas

Provider model availability and usage quotas depend on the configured account.

Run locally

git clone https://github.com/mtaimoor1407/DocuLegal.git
cd DocuLegal
python -m venv .venv

Activate with .venv\Scripts\activate on Windows or source .venv/bin/activate on macOS/Linux.

pip install -r requirements.txt

Create .env in the repository root:

GROQ_API_KEY=your_groq_api_key

Start the application:

streamlit run app.py

Open the local URL printed by Streamlit. The first run may download the embedding and reranking models.

Example workflow

Upload a document, index it, and ask a question such as:

• What notice period does this agreement specify?
• Which clause describes confidentiality obligations?
• What are the payment deadlines and stated late-payment consequences?
• How do the termination clauses differ between these two documents?

Check the cited passage in the original PDF. These are example prompts, rather than a claim that every uploaded document contains the requested information.

Repository guide

Path	Purpose
app.py	Streamlit interface
core/	PDF loading, chunking, retrieval, generation, and prompts
models/schemas.py	Structured response definitions
utils/	Supporting display/utilities
requirements.txt	Python dependencies
.streamlit/	Streamlit configuration

Evaluation and limitations

The confidence label is an application output, rather than a validated probability that an answer is correct. Source references can also be incomplete or incorrect and should be checked.

PDF extraction quality, missing context, retrieval errors, and ambiguous clauses can affect responses. Document comparison depends on retrieving the relevant evidence from each document.

A reproducible evaluation should include answerable and unanswerable questions, expected supporting passages, citation checks, retrieval Recall@k, answer correctness, and latency. Compare keyword-only, vector-only, hybrid, and reranked retrieval before attributing improvements to a specific technique. Quantitative results should be reported after those experiments are run.

DocuLegal is an informational prototype and does not replace professional legal advice.

Author: Muhammad Taimoor — BS Computer Science, COMSATS University Islamabad, Lahore Campus.