How does document chat work?

How document chat works

Vector Embeddings

Vector embeddings are a central part of our document chat system. They allow us to convert text parts from your document into numerical representations that can be analyzed efficiently. This is key to finding the most relevant parts of your document when you ask questions. Read more about this on Vector calculations.

Accurate answers based on your document

Our system uses the most relevant parts of your document to generate answers to your questions. This ensures that the answers are accurate and directly based on the content of your document, reducing the risk of inaccurate or irrelevant responses.

1. Document upload and chunking: Your uploaded document is divided into smaller chunks. Each chunk is embedded, meaning they are converted into numerical representations (embeddings) and placed on a unit circle. This allows the system to analyze and process data more efficiently.

How document chat works

2. Reformulating questions: When you ask a question about the document, a language model (LLM) reformulates the question into three different variations to extract relevant information from the indexed document. This ensures that all possible relevant answers are considered.

How document chat works

3. Retrieving relevant data: The most relevant chunks from your document are retrieved and re-embedded to ensure that the most precise and useful data is used to answer your question.

How document chat works

4. Generating the answer: The language model (LLM) combines your question with the retrieved information from the document and generates an answer that is sent back to you. This answer is based on both your original query and the relevant data found in your document, increasing the accuracy and relevance of the response.

How document chat works