Nexlit

Hey IH 👋

I'm a GenAI engineer and I built Nexlit (https://nexlit.co) — an AI-powered tool that lets researchers upload their PDF papers, then ask questions across all of them and get answers with exact page-level citations.

The problem I kept seeing:

PhD students and researchers have 50-200+ papers in their collection. When writing a literature review or thesis, they constantly need to find "which paper said what." The options are either re-reading papers for hours or doing ctrl+F keyword searches that miss anything phrased differently.

What Nexlit does:

1. Upload PDFs → auto-extracts text, generates summaries, indexes everything with semantic embeddings

2. Search by meaning, not keywords ("methods to reduce attention complexity" finds papers about FlashAttention, sparse attention, etc.)

3. Ask questions across your entire library → get structured answers with inline [1] [2] citations linking to the exact paper and page

Tech stack (for the curious):

• Frontend: React + TypeScript + Tailwind + Vite

• Backend: FastAPI + Python

• LLM: Claude Haiku 4.5 (for cost efficiency)

• Embeddings: OpenAI text-embedding-3-large

• Vector DB: Pinecone

• Database: Supabase PostgreSQL

• Storage: AWS S3

• Hosting: EC2 + CloudFront

The architecture that took the most iteration:

I use a 2-call RAG pipeline. Call 1 generates the answer without citations (so the LLM focuses purely on accuracy). Call 2 takes that answer + the retrieved chunks and injects accurate citations, verifying each claim against the source text. This separation improved citation accuracy significantly compared to asking the LLM to do both at once.

I also recently added hybrid search (BM25 + semantic) because pure semantic search misses exact terms like model names, metric numbers, and abbreviations.

I benchmarked it honestly:

I ran a RAGAS evaluation on 30 AI/ML papers with 20 expert-written questions. Results:

• Faithfulness: 93% (low hallucination)

• Response Relevancy: 98%

• Context Recall: 93%

• Factual Correctness: 62-66% (weakest — mostly a measurement artifact from F1 token overlap, but I published the real numbers including where it fails)

Full benchmark writeup: https://nexlit.co/blog/benchmark-nexlit-rag-accuracy-30-papers-ragas

Current status:

• Live at https://nexlit.co

• Free tier (no credit card needed)

• Solo founder, bootstrapped

• Ranking on Google page 1 for some long-tail research queries after ~1 month of SEO work

Would love feedback from anyone who works with research papers. What would make you switch from your current workflow?