AusLaw MCP

# Feature Development Roadmap ## Current State Analysis ### What Works Well ✅ - Fetches recent case law from AustLII - Filters out journal articles (primary sources only) - Extracts neutral citations and jurisdictions - Preserves paragraph numbers in `[N]` format (402 instances found in test) - Handles HTML and PDF documents ### Current Limitations 🔴 #### 1. **Search Quality Issues** - **Problem**: Searching "Donoghue v Stevenson" returns recent 2025 cases that merely cite it, NOT the actual 1932 case - **Root cause**: Sorting by date prioritises recent cases over relevance - **Impact**: Users can't find the specific case they're looking for #### 2. **Limited Sources** - **Current**: Only searches AustLII - **Missing**: jade.io (superior reported judgments), BarNet Jade, other authoritative sources - **Impact**: May miss best/most authoritative version of judgments #### 3. **Paragraph Number Preservation** - **Current**: Text extraction strips HTML structure - **Found**: `[N]` format markers ARE preserved (402 instances) - **Issue**: Page numbers from reported judgments are lost - **Impact**: Can't generate accurate pinpoints for reported citations #### 4. **No Ranking/Relevance** - **Problem**: No way to prioritise authoritative sources - **Missing**: Reported vs unreported distinction, court hierarchy weighting ## Proposed Solutions ### Phase 1: Fix Search Relevance (HIGH PRIORITY) **Goal**: Return the ACTUAL case being searched for, not just cases that cite it **Implementation**: 1. **Add search mode parameter**: `relevance` vs `date` sorting 2. **Smart query detection**: - If query looks like case name (e.g. "X v Y"), use relevance - If query is topic (e.g. "negligence"), use date for recency 3. **Title matching boost**: Prioritise exact title matches 4. **Citation matching**: Parse citations from query and match **Code changes**: ```typescript interface SearchOptions { jurisdiction?: "cth" | "vic" | "federal" | "other"; limit?: number; type: "case" | "legislation"; sortBy?: "relevance" | "date" | "auto"; // NEW } ``` ### Phase 2: Multi-Source Integration (MEDIUM PRIORITY) **Goal**: Search multiple authoritative sources and return best results **Sources to integrate**: 1. **jade.io** - Superior reported judgments with page numbers 2. **BarNet Jade** - Free access to some reported cases 3. **AustLII** - Comprehensive unreported coverage (current) **Implementation approach**: ```typescript // Parallel search across sources const [austliiResults, jadeResults] = await Promise.all([ searchAustLii(query, options), searchJade(query, options), // NEW ]); // Merge and deduplicate by citation const merged = deduplicateResults([...austliiResults, ...jadeResults]); // Rank by authority: Reported > Unreported, Higher court > Lower court const ranked = rankByAuthority(merged); ``` **Challenges**: - jade.io may require authentication/API key - Need to handle different HTML structures per source - Deduplication logic must match same case across sources ### Phase 3: Enhanced Paragraph/Page Preservation (HIGH PRIORITY) **Goal**: Preserve both paragraph numbers AND page numbers for accurate pinpoint citations **Current state**: - `[N]` paragraph markers: ✅ Preserved (402 found) - Page numbers: ❌ Lost in text extraction **Implementation**: 1. **Improve HTML parsing** to preserve structural markers: ```typescript // Keep paragraph markers <p class="Judg-Para-1">[1]</p> → "[1]" // Extract page markers (when present in reported versions) <span class="page-num">123</span> → "[Page 123]" ``` 2. **Return structured content**: ```typescript interface EnhancedFetchResponse extends FetchResponse { paragraphs?: Array<{ number: number; text: string; pageNumber?: number; }>; } ``` 3. **Pinpoint generation helper**: ```typescript function generatePinpoint( text: string, searchPhrase: string ): { paragraph?: number; page?: number } { // Find paragraph/page containing phrase } ``` ### Phase 4: Intelligent Result Ranking (MEDIUM PRIORITY) **Goal**: Return best/most authoritative version of each case **Ranking criteria** (in order): 1. **Reported vs unreported**: Reported judgments rank higher 2. **Court hierarchy**: HCA > Full Court > Single judge 3. **Completeness**: Judgments with page numbers > without 4. **Recency**: For topic searches, prefer recent 5. **Relevance**: Title/citation exact match > partial match **Implementation**: ```typescript function calculateAuthorityScore(result: SearchResult): number { let score = 0; // Reported judgment if (result.citation && !result.neutralCitation) score += 100; // Court hierarchy if (result.url.includes('/HCA/')) score += 50; else if (result.url.includes('/FCA/')) score += 30; // ... etc // Has page numbers if (result.metadata?.hasPageNumbers) score += 20; return score; } ``` ## Implementation Status ### ✅ Phase 1: Search Relevance (COMPLETED) **Implemented features:** 1. ✅ Smart query detection: Auto-detects case names ("X v Y", "Re X", citations) vs topic searches 2. ✅ `sortBy` parameter: "auto" (default), "relevance", or "date" modes 3. ✅ Title matching boost: Prioritizes exact case name matches in results 4. ✅ Auto mode intelligence: - Case name queries → relevance sorting to find specific cases - Topic queries → date sorting for recent jurisprudence 5. ✅ Comprehensive test suite: 7 new tests covering all sorting scenarios **What was fixed:** - ❌ **OLD**: Searching "Donoghue v Stevenson" returned 2025 cases citing it - ✅ **NEW**: Search returns the actual case being searched for **Technical details:** - Pattern detection for "X v Y", "Re X", citations, and quoted queries - Title scoring algorithm with party name matching - Configurable sorting with sensible defaults ## Implementation Priority ### Must Have (Next Sprint) 1. ✅ ~~Fix search relevance for case name queries~~ (COMPLETED) 2. ✅ ~~Preserve paragraph numbers properly~~ (already working) 3. ✅ ~~Add search mode parameter (relevance/date/auto)~~ (COMPLETED) ### Should Have (Following Sprint) 1. 🔶 Add jade.io integration for reported judgments 2. 🔶 Implement page number extraction 3. 🔶 Add authority-based ranking ### Nice to Have (Future) 1. 📋 BarNet Jade integration 2. 📋 Citation parsing and validation 3. 📋 Automatic pinpoint generation 4. 📋 Related cases/legislation suggestions ## Testing Requirements Each phase must include: 1. **Unit tests** for new parsing/ranking logic 2. **Integration tests** with real judgments 3. **Comparison tests** - verify improvements over current state 4. **Performance tests** - ensure multi-source doesn't timeout ## Success Metrics 1. **Search accuracy**: "Donoghue v Stevenson" returns the 1932 case, not 2025 cases 2. **Pinpoint accuracy**: Can generate `[2025] HCA 26 at [42]` style citations 3. **Source coverage**: Returns reported judgment when available 4. **Response time**: < 5 seconds for multi-source search