Chapter 13 — Known Edge Cases & Future Roadmap

Edge Case	Description	Current Impact	Proposed Fix
Seventh Schedule Overlap	Schedule entries use numbered lists ("19. Price control", "34. Betting"). The chunker's regex tags these as article_number: "19".	Low — LLM correctly differentiates in its response but adds an unnecessary note	Enhance regex to verify chunk resides within Part I–XXII before tagging
General Conceptual Queries	Queries like "What are Fundamental Rights?" don't specify an article number, so metadata filter is not applied.	Medium — relies on semantic search, which is good but not 100% precise	Add part metadata (e.g., part: "III") to enable Part-level filtering
Cross-Article References	Some articles reference others (e.g., Article 32 protects rights under Article 19).	Low — system answers correctly for individual articles but doesn't auto-link related ones	Future: build a Knowledge Graph (Neo4j) to map inter-article relationships

Hallucination-Resistant RAG — Constitution of India Case StudyPage 20

Chapter 14 — Key Takeaways (Interview-Ready)

Interview Prep

If asked: "How did you handle RAG for a complex legal document?"

The Architecture Pattern

Layer	Technique	Industry Term
Parsing	PyMuPDF + Custom Regex Cleaning	Structure-Aware Document Parsing
Chunking	Article-boundary splits with parent → child hierarchy	Hierarchical Chunking / Multi-Granularity Chunking
Metadata	Dynamic Article number extraction & injection	Metadata-Enriched Vector Indexing
Retrieval	LLM Router → Metadata Filter → Pinecone	Hybrid Retrieval with Metadata Filtering
Overall	8-Node LangGraph Pipeline with classification, retrieval, generation, hallucination guard	Agentic RAG Architecture

The One-Liner

"I replaced expensive LLM parsing with a deterministic, structure-aware pipeline using Hierarchical Chunking and strict metadata filtering — achieving hallucination-resistant retrieval on a 400-page legal document at zero parsing cost."

— End of Case Study — © Ambuj Kumar Tripathi | All Rights Reserved