tbroadley / ai-safety-conversational-agent

AI Safety Conversational Agent

This is a fork of chat-langchain. It takes data from https://aisafety.info and uses it to create a chatbot that can answer questions about AI safety.

✅ Running locally

1. Install dependencies: pip install -r requirements.txt
2. Download this folder and extract it into a folder called aisafety.info
3. Run ingest.sh to ingest the data in aisafety.info
4. Run the app: make start
   1. To enable tracing, make sure langchain-server is running locally and pass tracing=True to get_chain in main.py. You can find more documentation here.
5. Open localhost:9000 in your browser.

📚 Technical description

There are two components: ingestion and question-answering.

Ingestion has the following steps:

1. Pull documents from Google Drive
2. Load documents with LangChain's document loaders
3. Split documents with LangChain's TextSplitter
4. Create a vectorstore of embeddings, using LangChain's vectorstore wrapper (with OpenAI's embeddings and FAISS vectorstore).

Question-Answering has the following steps, all handled by ChatVectorDBChain:

1. Given the chat history and new user input, determine what a standalone question would be (using GPT-3).
2. Given that standalone question, look up relevant documents from the vectorstore.
3. Pass the standalone question and relevant documents to GPT-3 to generate a final answer.

TODOs

• Train on https://github.com/moirage/alignment-research-dataset