Lexis legal intelligence

Inspiration

Every business faces legal questions daily. An employee complaint lands in your HR queue. A customer dispute needs resolution. A compliance question blocks a project.

Your team creates a Jira ticket... then what? Google returns blog posts, not legislation. ChatGPT confidently cites laws that don't exist. Lawyers cost $300/hour and take days to respond.

We built Lexis because legal questions shouldn't require a law degree to answer.

What it does

Lexis Legal Intelligence brings New Zealand legislation directly into Atlassian:

🔍 Jira Issue Panel - Open any issue and instantly see relevant legislation sections. Employment dispute? Here's the Employment Relations Act sections that apply.

✅ Grounded Citations - Every claim cites a real section number. Click any citation to open the official legislation.govt.nz page. No hallucinations.

💬 Legal Chat - Click "Continue in Chat" for deeper analysis. Ask follow-up questions, get cited answers.

📄 Confluence Macro - Embed legal research in documentation with the /lexis macro.

How we built it

Backend Architecture:

• Flask API hosted on Railway
• Genesis holographic semantic search (sub-3 second retrieval)
• Structured extraction with GPT-4o-mini for grounded responses
• Domain discovery algorithm (detects legal domain from search results, not prediction)

Corpus:

• 4000 NZ legislation sections
• 19 legal domains (employment, consumer, tenancy, privacy, corporate, tax, and more)
• Every section linked to official legislation.govt.nz URL

Frontend:

• Forge UI Kit 2 with @forge/react
• DynamicTable for sortable citations
• ADF renderer for structured legal text

Key Innovation - Grounded AI: Unlike ChatGPT which generates plausible-sounding law, Lexis uses a two-stage approach:

1. Retrieve real legislation using semantic search
2. Extract facts as structured JSON (the AI copies, it doesn't create)
3. Render with mandatory citations (code, not AI)

Result: 95%+ grounded responses where every claim traces to source.

Challenges we faced

Challenge 1: Relevance Early versions returned irrelevant legislation (H&S sections for employment queries). We solved this by implementing "domain discovery" - analyze search results to find the dominant legal domain, then filter.

Challenge 2: Speed Legal briefs took 40+ seconds. We optimized to sub-3 second search with async extraction, making it usable in real workflows.

Challenge 3: Grounding Preventing AI hallucination in legal context is critical. We switched from "write a brief" (freeform) to "extract facts as JSON" (constrained), dramatically improving accuracy.

What we learned

• Discovery beats prediction - Let results tell you the domain, don't guess beforehand
• Structured extraction > freeform generation - Constrain the AI to copy, not create
• Citations build trust - One verifiable link is worth more than paragraphs of explanation

What's next

• 🇦🇺 Australian legislation corpus
• ⚖️ Case law integration
• 🏢 Enterprise features for legal teams
• 🔄 Jira Automation triggers for compliance workflows ```

---

Built With

• atlassian
• docker
• flask
• forge
• openai
• postgresql
• python
• pytorch
• railway
• react