joyboseroy/falkor-irac

falkor-irac is a reasoning framework for Indian legal AI that uses a knowledge graph to verify every answer an AI generates before returning it to the user. It is built around the concept that legal reasoning is not fuzzy text search, it is structured traversal of relationships between cases, statutes, and precedents. The system ingests Supreme Court and High Court judgments and stores them as structured nodes in a graph database called FalkorDB. Each judgment is broken into IRAC components, Issue, Rule, Analysis, Conclusion, and connected to related cases, statutes, judges, procedural events, and outcomes through typed relationships. The README describes this as treating courts as "graph traversal engines disguised as prose." When a user asks a legal question, a retrieval agent traverses the graph to find relevant precedents and reasoning paths. An LLM then generates an answer guided by those paths. Before the answer is returned, a Verifier Agent checks whether a valid citation path actually exists in the graph supporting the proposed answer. If no such path exists, the answer is rejected or flagged. This binary check, the README calls it a "falsifiability oracle", is the core anti-hallucination mechanism. The system also explicitly detects doctrinal conflicts: when two valid paths support contradictory conclusions, it returns both paths, labels the conflict type (such as coordinate bench disagreement or overruled precedent), and flags it for human review rather than silently choosing one answer. It is written in Python and uses FalkorDB as the graph database, run via Docker.