HN Showcase
Back to browse
GitHub Repository

LexReviewer is an AI legal document review engine that ingests PDFs, uses RAG for hybrid retrieval, and enables grounded chat with citations and bounding-box references, delivering verifiable answers directly from the source text.

11 starsPython

LexReviewer – Because "Chat with PDF" is broken for legal workflows

by sherebanuk·Feb 24, 2026·1 point·1 comment
Visit ProjectView on HN

AI Analysis

●●SolidSolve My ProblemBig Brain

LangGraph agent adapts search strategy per query, but LLMs still hallucinate in contracts.

Strengths
  • •LangGraph-based adaptive retrieval strategically chooses search mode (semantic vs keyword vs linked docs) instead of one-size-fits-all RAG
  • •Citation-aware streaming with bounding boxes grounds answers in source text, solving real legal verification friction
  • •Multi-document awareness follows references (amendments, schedules, MSAs) across related PDFs in one context
Weaknesses
  • •Relies on OpenAI API — no offline or fine-tuned model option for legal domain where accuracy is existential
  • •Architecture is orthodox RAG+agent orchestration; no novel technique beyond applying LangGraph to a well-understood problem
Category
AI/ML
Target Audience

Legal professionals, contract reviewers, compliance teams

Similar To

LlamaIndex (formerly GPT Index) · Langchain + FastAPI RAG systems · Perplexity (retrieval-augmented generation)

Post Description

Hi HN!

Most “chat with PDF” tools work fine until you try using them for something that actually matters, like contracts.

The issue isn’t that they can’t answer questions. It’s that you can’t trust the answers. They return something that sounds correct, but don’t clearly show where it came from, or they miss context from referenced clauses and related documents.

Legal docs make this harder because questions aren’t uniform: - sometimes you’re searching concepts - sometimes exact clause IDs - sometimes text from a different linked document

Most systems handle all of those the same way, which is where things break.

So I built LexReviewer, an open-source backend designed around a single rule: ""an answer isn’t useful unless you can verify it instantly.""

Instead of treating every query identically, it adapts its search strategy based on what you’re asking and can follow references across documents when needed. The result is answers that stay grounded in real text and point directly to the source passage.

Repo: https://github.com/LexStack-AI/LexReviewer

-- Currently tested on 300+ page contracts with cross-references

Feedback I’d especially value:

- Where do current document-AI systems fail hardest for you? - What’s been the biggest blocker to trusting AI outputs in production workflows? - If you’ve built something similar, what design choices ended up mattering most?

Similar Projects