Large language models like ChatGPT face challenges with token limits and incorporating private data, which restricts their use on large documents or custom knowledge bases.
Unlike relational or NoSQL databases, vector databases are designed to store and semantically search unstructured data, filling a critical gap in the data landscape.
Manually tagging unstructured data like images and documents is inconsistent and subjective, making it an inefficient way to enable search.
Machine learning models convert unstructured data into numerical representations called embeddings, where semantically similar items are positioned closely in a high-dimensional space.
A demonstration with Google's Projector TensorFlow shows how words like "king" and "queen" are clustered together, visually representing their semantic proximity.
Vector databases use mathematical formulas to measure the distance between embeddings and employ indexing techniques like Approximate Nearest Neighbor (ANN) for high-speed search.
A look at popular vector databases like Pinecone, Weaviate, and Milvus, including their features, hosting models, and integrations with platforms like Hugging Face.
Vector databases can power real-world applications such as intrusion detection systems and face similarity matching without needing constant model retraining.
The Retrieval-Augmented Generation (RAG) pattern uses a vector database to find relevant data chunks, providing LLMs with the right context to answer questions accurately.
Vector search enables a wide range of applications including recommendation systems, document deduplication, time-series analysis, and advanced product search.
58:00Vijay Krishan Gupta & Gauravdeep Singh Lotey
57:52Rainer Stropek
56:55Ekaterina Sirazitdinova
59:39Julian Joseph
27:23Nico Martin
1:00:56Mary Grygleski
47:28Tanmay Bakshi
26:55Jan Zawadzki
Jobs that call for the skills explored in this talk.
Deutsches Medizinrechenzentrum GmbH
DTS Systeme GmbH
LEONHARD WEISS GmbH & Co. KG
Understanding Recruitment Group
Barcelona, Spain
