LegalGPT is Claw LegalTech’s proprietary Generative AI assistant tailored for Indian legal professionals. The tool allows users to perform contextual legal research, draft contracts, and understand court judgments by interacting with a chat-based interface powered by a custom-trained LLM.

At its core, LegalGPT is a hybrid RAG (Retrieval-Augmented Generation) system layered with multiple legal-specific capabilities, including citation insertion, clause extraction, judgment summarization, and precedent suggestion.

To build a domain-specific LLM system optimized for legal texts with reliable, up-to-date Indian case law and legislation. The system needed a scalable and maintainable ETL pipeline to handle continuously evolving legal corpora.

Case Laws: Extracted from Indian judiciary portals (e.g., Indiankanoon, eCourts, Judis)

Statutes and Bare Acts: Scraped from government websites in structured and unstructured formats

Legal Contracts: Anonymized document bank from law firms (PDFs, DOCX, HTML)

Law Firm Blogs and Treatises: Extracted with semantic weighting for relevance

Built using Scrapy + Playwright for dynamic scraping

Used asyncio-based job runners to schedule crawls for over 200k+ case law documents per month

Integrated pdfminer and docx parsers for contract ingestion

Data Cleaning: Removal of procedural text, pagination errors, footnotes

Clause and Entity Tagging: Custom spaCy NER pipeline to tag:

Semantic Chunking: Each case split into:

Stored as structured JSON blobs with metadata

Used HuggingFace Sentence-BERT (trained on legal domain data) for vector generation

Indexed embeddings into:

Embedding refresh jobs scheduled via Airflow

Metadata stored in PostgreSQL

Vector data in Pinecone

Raw + transformed data backed up in AWS S3 with versioning

Base Model: Mistral 7B / Falcon 7B fine-tuned with QLoRA on legal-specific instruction-tuning tasks

Integrated with LangChain to support:

Implemented context window trimming with token-weighting heuristics

Real-time legal precedent suggestions based on chat queries

Clause-level case law lookup and citation injection

Auto-tagging and classification of uploaded judgments

Legal document QA on contracts with “hallucination prevention” via RAG fallback

Data volume: Over 2M+ case docs — resolved with asynchronous batching and delta-updates using checksum logic

Document variability: Built rule-based + ML hybrid pipeline for structure detection

Vector index sync: Automated invalidation and refresh pipelines with Faiss ID hash matching

LegalGPT reduced legal research time by 80%

Supported drafting of legal arguments with jurisdiction-aware citations

The system is currently used by over 300 lawyers across India in alpha and pilot testing