Forager‑Lite

Inspiration

Honestly, this whole project started because we kept getting hit with random “facts” online that sounded real… until they weren’t. In school, in group chats, even in homework, people throw around claims like they’re 100% true. And most of the time, nobody checks anything.

We wanted to build something that makes fact‑checking easy enough that people actually do it. Not a giant AI model, not a complicated research tool — just a simple agent that grabs evidence, shows you where it came from, and tells you how confident it is.

That idea turned into Forager‑Lite, our attempt to make truth‑checking feel less like homework and more like a normal part of using the internet.

What it does

Forager‑Lite takes any claim like “The Battle of Hastings happened in 1066” — and automatically:

searches reliable sources

pulls out the most relevant snippets

compares them to the claim

scores how similar they are

shows you the evidence

gives you a confidence score

It’s basically a tiny verification agent that doesn’t try to “think” for you, it just shows you the receipts.

How we built it

We built a FastAPI backend that exposes a clean /verify/claim endpoint. It connects to:

the Wikipedia API

Bing Web Search (if you add a key)

a fuzzy‑matching system using rapidfuzz

a weighting system that ranks evidence

a confidence calculator

Then we built a Streamlit frontend so anyone can paste a claim and instantly see the results.

We deployed the backend on Render using Docker and the frontend on Vercel. Everything is environment‑variable based, so no keys ever get committed.

We basically built a full agent pipeline without needing any heavy AI models.

Challenges we ran into

We’re teenagers, so we’re used to debugging homework at 1AM — but this project gave us some real challenges:

Wikipedia snippets are messy and unpredictable

Bing results sometimes return weird formatting

fuzzy matching can be too strict or too forgiving

CORS between Render and Vercel fought us for days

we had to design a scoring system that actually makes sense

we had to make sure nothing ever hallucinated or made up facts

We also had to learn how to build something that feels trustworthy. That meant showing evidence clearly, explaining confidence scores, and making sure the agent never pretends to “know” anything.

Accomplishments that we're proud of

We’re proud that we built something that actually works — not just a demo, but a real verification agent that:

is fully deterministic

shows every piece of evidence

explains exactly why it scored something the way it did

doesn’t hallucinate

doesn’t hide its reasoning

is fast enough for real use

is simple enough that anyone can understand it in under 30 seconds

We’re also proud that we built this as students. We didn’t have fancy GPUs or huge models — just logic, APIs, and a lot of debugging.

What we learned

We learned how to:

build an agentic system without relying on AI

design a pipeline that’s transparent and safe

deploy microservices in the real world

handle messy external APIs

write code that other people can actually use

build something that feels trustworthy

We also learned that “agentic AI” doesn’t always mean “big AI model.” Sometimes it just means building a system that can take a task, break it down, and do it safely.

What's next for Forager‑Lite

We want to take Forager‑Lite further:

add semantic embeddings for deeper similarity

add more evidence sources (academic, .edu, .gov)

build a dashboard for batch verification

add caching and rate‑limiting

improve the UI

add a “source reliability” score

maybe even build a Chrome extension

Our goal is to make fact‑checking something people actually do — not because they’re forced to, but because it’s finally easy.

Built With

• dockerfile
• eml
• fastapi
• javascript
• json
• llmextraction
• markdown
• python
• regex
• shell
• sql
• uvicorn