/fakt_ai

Fully transparent, automated fact checking

Primary LanguageJupyter Notebook

🔍 Fakt.ai

Fully transparent, automated fact-checking powered by AI Agents

Try it for free here: https://faktai.streamlit.app/

⚠️ The Problem

  • Have you ever read something online and thought "I don’t know if that’s true"?
  • Have you ever heard someone confidently speaking about a tense topic? They quote research papers saying you can "look it up" and "read them yourself".
  • Have you ever actually tried to look them up?

I did. And hit two major obstacles:

  1. Research is hard to find.
  2. Research is hard to understand. Does a paper saying they observed a 6% uptake in
    protein R-X-517A across all participants support what the original speaker was saying?

✅ The Solution

We need a tool that:

  • Finds research papers that support/contradict given arguments
  • Analyses said papers for quality (e.g. are they biased lobbying groups, is the sample size too small?)
  • Compares and contrasts arguments to conclude whether the original claim is true
  • Is fully transparent, so the user can see each step of the way

Enter Fakt AI - fully transparent, automated fact-checking powered by AI Agents.

⚙️ Installation

  1. Install Poetry

    curl -sSL https://install.python-poetry.org | python3 -

  2. cd into the project folder and install dependencies with

    poetry install

  3. Create an .env file and add the follwing API keys

    GROQ_API_KEY=...              (required)
SEMANTIC_SCHOLAR_API_KEY=...  (optional, results in more successful runs)
    • Grok: Sign up to Grok Cloud and create an API key
    • Semantic Scholar: follow the instructions here (it will take a few days to arrive, annoyingly)

  4. Run a query (takes 1-2 minutes)

    poetry run python main.py --query "were treatments available to treat COVID before the vaccines came out?"

Note: at the moment, Fakt AI only supports academic paper search. So please ask questions that require academic papers to answer them.

Note 2: it can be a bit tempermental. If you do not get back an answer, please re-run the query. Usually it works the second time.

💪 Areas for Improvement / Roadmap

  • Deploy as a front-end for nicer user interactions
  • Clickable URL links for every reference
  • Support for more than just academic paper search
  • Run analysis on every reference (e.g. bias likelihood check, analyse any experiments in the paper, is the sample size too small etc.)
  • Chat with the agent after e.g. ask questions about specific bits of the process or to search more in certain areas
  • Query optimisation/breakdown. Breakdown queries into multiple steps and perform multiple searches to improve results
  • Faster!
  • A nice UI to be able to step back through the thinking and easily open up papers and see relevant sections
  • Handle questions that even the experts have not come to a conclusion about e.g. 60% of the data supports this PoV, 40% supports the opposite.

💰 Monetisation

  • Pay per fact check - powered by the FaktAI token

Who?

  • Journalists
  • Academics
  • Researchers
  • General people who want to be better informed / know the truth e.g. watch Russel Brand, use X, follow Robert F Kennedy.