CaseVault GraphRAG
Privacy-first legal research workspace that ingests case documents, ranks search results by relevance, and provides AI-generated summaries with verifiable citations.
[ 01 ] — Context
Problem
- Bangladeshi law firms work with thousands of handwritten and scanned documents, losing annotations and struggling to find relevant precedents.
- Generic AI chatbots hallucinate citations — and firms cannot upload confidential client files to public AI tools without risking privacy.
[ 02 ] — Response
Solution
CaseVault Phase 1 is a working MVP: a FastAPI backend ingests markdown legal documents (front-matter metadata) into SQLite, with repositories that score document relevance per query and a full-text search API.
The Next.js frontend is a dark-theme research workspace — hero search, live stats, category cards, a document reader with AI tabs (summary / ask / citations / related), query highlighting, and an XSS-safe hand-rolled markdown renderer.
Phase 2 is fully designed (Project SynthGraph): Qdrant semantic search + Neo4j knowledge graph with Leiden community detection, Celery pipelines for OCR → chunking → embeddings, and sub-graph context injected into a reranker — targeting ~790ms to first token.
[ 03 ] — Structure
Architecture
┌───────────────────────── Phase 1 (shipped) ─────────────────────────┐ │ Next.js 16 SPA ──► FastAPI ──► SQLAlchemy ──► SQLite │ │ / (search) /reader /admin /dashboard /library /workspace │ │ DocumentService.sync_documents_from_disk() → 46 legal docs │ └──────────────────────────────────────────────────────────────────────┘ ┌───────────────────────── Phase 2 (designed) ────────────────────────┐ │ OCR → chunk → embed → Qdrant (vector) · Neo4j (KG, Leiden) │ │ Celery async pipelines · BGE reranker · ~790ms TTFT budget │ └──────────────────────────────────────────────────────────────────────┘
[ 04 ] — Trade-offs
Key engineering decisions
- 01
XSS-safe markdown without a heavy renderer
The reader renderer is hand-rolled and output-sanitized rather than pulling in a full markdown engine — small surface, safe by default for legal content.
- 02
Relevance scoring at the repository layer
Document relevance is computed in SQL against indexed columns and cached, keeping the search API fast without introducing a search server in Phase 1.
- 03
Phase-2 placeholders designed up front
Routes like /admin, /chat, and /workspace exist as intentional shells so the product structure survives the Phase 1 → Phase 2 migration.
[ 05 ] — Under the hood
Engineering highlights
def search(self, q: str, limit: int = 20) -> list[Document]:
like = f"%{escape_like(q)}%"
rows = self.session.query(Document).filter(
or_(Document.title.ilike(like),
Document.content.ilike(like))
).order_by(relevance_rank(q)).limit(limit).all()
return rows[ 06 ] — Outcomes
Results
[ 08 ] — Where it's going
Status & next steps
- Phase 2 build: Qdrant + Neo4j + Celery pipelines per the SynthGraph design.
- Admin review queue where firms approve AI-suggested citations before they reach research notes.
[ 09 ] — Keep exploring
More projects
InvoicePilot
AI-powered invoice processing platform that extracts, validates, and analyzes invoices while detecting duplicates and anomalies to streamline financial document workflows.
LedgerTurf
Real-time turf booking platform that lets players discover and reserve grounds on a map while owners publish and manage slots with overlapping-time protection.