Build RAG Australian Legal AI Assistant
Budget / SalaryHourly project
TypeFreelance project
LocationRemote
Posted1 hour ago
I’m building a retrieval-augmented AI assistant that confidently answers questions about Australian criminal, civil and corporate law. The system must support three core workflows: precise legal research, conversational Q&A and first-draft document generation. Every answer has to cite real case law, legislation and reputable legal articles—no hallucinations.
The job centres on designing and implementing the full RAG pipeline:
• Data ingestion & preprocessing – scrape or load authorised reports, consolidated legislation and selected secondary commentary, then chunk, clean and embed.
• Retrieval layer – stand up a performant vector store (FAISS, Pinecone or similar) with filtering for jurisdiction, court level and date.
• LLM orchestration – wire the retriever to an LLM of your choice (open-source or API-based) with custom prompting and answer-structuring that keeps citations intact.
• Evaluation – build automatic tests measuring factual accuracy, citation integrity and latency; refine until zero-hallucination rates are met on a held-out set.
• Deployment – containerise the service, expose a REST/GraphQL endpoint and supply a short demo notebook or script performing legal research, Q&A and sample document drafting.
Acceptance criteria
1. Given a user query, the assistant returns a concise answer plus clickable citations pointing to the underlying case law or statute section.
2. Retrieval accuracy ≥ 90 % on a sampled validation set; hallucination rate ≤ 2 %.
3. End-to-end latency < 5 s for typical queries on a mid-range GPU or CPU beat.
4. Codebase delivered with clear README and environment files so another engineer can reproduce results in one command.
If this aligns with your expertise in modern ML (Python, LangChain/LlamaIndex, vector databases, containerisation) and you have experience working with legal or highly regulated text, I look forward to seeing how you would approach the build and timeline.
The job centres on designing and implementing the full RAG pipeline:
• Data ingestion & preprocessing – scrape or load authorised reports, consolidated legislation and selected secondary commentary, then chunk, clean and embed.
• Retrieval layer – stand up a performant vector store (FAISS, Pinecone or similar) with filtering for jurisdiction, court level and date.
• LLM orchestration – wire the retriever to an LLM of your choice (open-source or API-based) with custom prompting and answer-structuring that keeps citations intact.
• Evaluation – build automatic tests measuring factual accuracy, citation integrity and latency; refine until zero-hallucination rates are met on a held-out set.
• Deployment – containerise the service, expose a REST/GraphQL endpoint and supply a short demo notebook or script performing legal research, Q&A and sample document drafting.
Acceptance criteria
1. Given a user query, the assistant returns a concise answer plus clickable citations pointing to the underlying case law or statute section.
2. Retrieval accuracy ≥ 90 % on a sampled validation set; hallucination rate ≤ 2 %.
3. End-to-end latency < 5 s for typical queries on a mid-range GPU or CPU beat.
4. Codebase delivered with clear README and environment files so another engineer can reproduce results in one command.
If this aligns with your expertise in modern ML (Python, LangChain/LlamaIndex, vector databases, containerisation) and you have experience working with legal or highly regulated text, I look forward to seeing how you would approach the build and timeline.
Apply on Freelancer →
Project sourced from Freelancer.com. Applications happen directly on the original platform — we never collect your data.