Analyze legal contract risk with Google Gemini hybrid RAG and Supabase

🚀 What This Workflow Does This workflow transforms any PDF legal contract into a detailed AI-powered risk report — in under 5 minutes. Upload a contract, and the system automatically splits it into clauses, analyses each one using Hybrid RAG (semantic + keyword search), scores risk as HIGH / MEDIUM / LOW, and delivers plain-English explanations with safer alternative wording. --- 🔥 Why Hybrid RAG? Most dangerous clauses don't use obvious legal keywords. "The Client accepts full responsibility for all third-party claims" is an indemnification clause — but keyword search misses it. Hybrid RAG combines: - Vector Search (pgvector) — finds semantically similar risky patterns - BM25 Keyword Search — catches explicit legal red flags - RRF Reranking — merges both results with clause-type boosting --- 🔍 What It Does - Accepts a PDF contract via webhook (with async jobid tracking) - Splits contract into individual numbered clauses - Classifies each clause type using Google Gemini (indemnification, IP, termination, etc.) - Generates vector embeddings and searches a Supabase knowledge base - Scores each clause HIGH / MEDIUM / LOW using regex + AI - AI Agent (Gemini Flash) explains risk in plain language + suggests safer wording - Aggregates all results into a single JSON report - Saves report to Supabase (frontend polls for result asynchronously) --- ⚙️ Architecture (Two Pipelines) Pipeline 1 — Ingestion: Builds the knowledge base of risky clause patterns in Supabase Pipeline 2 — Query: Analyses new contracts against the knowledge base Both pipelines run in the same workflow — the branch splits at Extract Embedding. --- 🧠 Key Technical Decisions - Async architecture — Frontend fires request + polls Supabase. No timeout issues. - jobid tracking — Preserved across all nodes via ...$json spread - RRF Reranking — Combines vector + BM25 scores with type-based boost multipliers - Regex Risk Scorer — First-pass risk classification before expensive LLM call - Gemini Flash — Fast, cost-efficient LLM for per-clause annotation --- 📦 Requirements - Google Gemini API key — for clause classification + embeddings + AI Agent - Supabase project — with pgvector extension enabled - Supabase tables: legalclauses (knowledge base) + reports (results) - Supabase functions: matchclauses() + keywordsearchclauses() - Frontend (optional): HTML/CSS/JS web app hosted on Netlify --- 💡 Example Use Cases - Freelancers reviewing client contracts before signing - Startups evaluating vendor or investor agreements - Legal ops teams standardising contract review at scale - Business owners catching risky clauses without legal fees --- 🎯 Output - Per-clause: risklevel, plain-English explanation, riskreason, saferalternative, keyobligations, legalarea - Summary: overallriskscore, riskdistribution, legalareas map, highriskclauses list - Stored as JSON in Supabase reports table, keyed by jobid