Embedding vectors convert complex concepts like text or personality into multi-dimensional numerical arrays, enabling comparison and clustering.
The OpenAI API provides an endpoint to generate a 1,536-dimensional vector for a given text, and vector similarity can be efficiently calculated using a dot product.
The chat completions API allows developers to build custom applications by sending a model the entire chat history, including system prompts and user messages.
The RAG pattern enhances LLM responses by first retrieving relevant facts from a private knowledge base using vector search and then injecting that context into the prompt.
A practical demonstration shows how to build a Q&A assistant for a school's private wiki using a crawler, an indexer, and a query application.
The first step in the RAG pipeline involves building a custom crawler to extract, clean, and convert source data into a usable format like Markdown.
An indexer application iterates through pre-processed documents, calculates their embeddings via the OpenAI API, and stores them in Azure Cognitive Search for fast retrieval.
The query application generates an embedding for the user's question, performs a vector search to find relevant documents, and injects them into a system prompt for the LLM.
The command-line assistant successfully answers specific questions about school policies by retrieving information from the wiki, even handling multi-language queries.
The speaker answers audience questions about how embeddings are calculated, ensuring answer correctness, responsible AI development, and recommended developer tools.
26:25Ashish Sharma
58:06Anna Fritsch-Weninger
56:55Ekaterina Sirazitdinova
59:38Erik Bamberg
1:09:49Simi Olabisi
24:24Damir
27:12Damir
31:39Rainer Stropek
Jobs that call for the skills explored in this talk.
Devoteam
DTS Systeme GmbH
