Knowledge Base LLM

Use a Vector Database to retrieve source knowledge from a Knowledge Base (Prolog) and let an LLM use the source to generate answers.

Create Knowledge Base

First we want to create a Knowledge Base (Prolog) that the LLM can reference for answering questions.

It also helps to

• Improve embedding accuracy for better results.
• Reduce text input to improve LLM efficiency.
• Provide precise information sources for users.
• Split long chunks of text that exceed context window.

Step by Step

1. Tokenize the text (e.g. with tiktoken).
2. Use langchains RecursiveCharacterTextSplitter to generate chunks of text
3. Create Embeddings for all chunks
4. Initialize a Vector Database and store the vectors in there
5. Use the RetrievalQA chain with a LLM and the Vector Database to generate answers for questions
6. We can also use the RetrievalQAWithSourcesChain to make the answers even more reliable.