AFLHR Lite

Adaptive Framework for LLM Hallucination Reduction — Lite Version

A two-layer verification pipeline that combines Retrieval-Augmented Generation (RAG) with Natural Language Inference (NLI) to detect hallucinations in LLM outputs. The core innovation is Confidence-Weighted CONLI (Cw-CONLI): the NLI verification threshold adapts dynamically based on retrieval confidence.

Quick Start

git clone https://github.com/shaunyogeshwaran/Shaun_FYP.git
cd Shaun_FYP
make install        # installs pip + npm dependencies, creates .env
# Edit .env and add your GROQ_API_KEY (free at https://console.groq.com)
make start          # starts backend (:8000) + frontend (:5173) + docs (:4000)

Open http://localhost:5173 to use the app.

Requirements

Requirement	Details
Python	3.10+ with pip
Node.js	18+ with npm
RAM	24 GB recommended (ML models load into memory)
Disk	~3 GB (models auto-download from HuggingFace on first run)
GPU	Optional — CUDA auto-detected, falls back to CPU
Groq API key	Optional — offline mode works without it

Make Targets

make start      # starts backend (8000) + frontend (5173) + docs (4000)
make stop       # stop all servers
make restart    # bounce all servers
make status     # check what's running
make install    # install all dependencies
make smoke      # smoke test (precompute 20 samples)

Frontend Pages

Page	Route	Description
Verify	/	Enter a claim, adjust thresholds, see the full pipeline and verdict
Explore	/explore	Batch-run 7 pre-configured queries across domains
How It Works	/about	Visual walkthrough of the 4-stage pipeline

Troubleshooting

Problem	Fix
make install fails on pip	Ensure Python 3.10+ is on your PATH
Backend won't start	Check logs: tail /tmp/aflhr_backend.log
Frontend won't start	Check logs: tail /tmp/aflhr_frontend.log
Port in use	make stop first, or kill the process on that port
Models downloading slowly	First run downloads ~3 GB from HuggingFace
MPS/Apple GPU segfault	Expected — system auto-disables MPS and uses CPU