Amazon Titan FMs
Amazon Titan AI
The 25 years of Amazon experience in artificial intelligence (AI) and machine learning (ML) developments form the foundation of the Amazon Titan family of models, which are only available in Amazon Bedrock. Through a fully managed API, Amazon Titan foundation models (FMs) provide an extensive library of pre-trained image, multimodal, and text models. Amazon Titan models are strong and adaptable, built for a variety of applications while abiding by responsible AI principles. They have been trained on large datasets.
Amazon Titan Text Embeddings V2, the company’s second-generation text embeddings model, is the newest member of the Amazon Titan family and is now accessible through Amazon Bedrock. The optimised text embeddings model for Retrieval-Augmented Generation (RAG) is this one. It has pre-trained on code and more than 100 languages.
Amazon Titan
Advantages
A wide variety of uses
Strong text, picture, and multimodal models can drive a wide range of generative AI applications, including recommendation and search functions, content production, and image generation.
Provide pertinent search results
Improve personalised recommendations and search accuracy.
Integrated assistance for conscientious AI
By screening model outputs, rejecting unsuitable user inputs, and identifying and eliminating harmful content from the data, all Amazon Titan FMs provide built-in support for the responsible use of AI. By default, every image produced by the Amazon Titan Image Generator model has an invisible watermark to assist prevent the creation of dangerous content and stop the dissemination of false information. One of the first major cloud providers to extensively offer built-in watermarks for AI picture outputs is AWS. Watermark detection is an innovative technology.
Simple personalisation
Utilise your own data to fine-tune Amazon Titan models so they may be customised to carry out tasks unique to your organisation.
Use cases
Text production
For a wide range of text-related operations, including writing material for web sites and blogs, categorising articles, conducting open-ended Q&A, conversational chat, information extraction, and more, use Titan Text models to increase productivity and efficiency.
Synopsis
To swiftly and efficiently extract crucial information from lengthy publications like articles, reports, research papers, technical documentation, and more, use Titan Text models to generate succinct summaries of the content.
Search semantics
To give end users more precise and contextually relevant multimodal search, recommendation, and personalisation experiences, leverage Titan Multimodal Embeddings and Titan Text Embeddings.
Creation of images
Give content producers the tools they need to quickly generate ideas and refine them into highly effective image creation. Using natural language cues, clients in the media and entertainment, e-commerce, and advertising sectors can produce realistic, studio-quality photographs in big quantities and at a reasonable cost.
Generation Augmented for Retrieval (RAG)
By linking FMs to your data sources, you may provide users with more accurate and up-to-date responses for their queries. Increase the Titan models’ already formidable powers and give them additional domain and organisation expertise.
Amazon Titan Model
Titan Text Express
LLM provides a performance-price balance.
Maximum tokens: 8,000
Languages: 100+ languages available (preview), English (GA)
Adjustment support: Indeed.
Retrieval augmented generation, open-ended text generation, brainstorming, summarization, code generation, table creation, data formatting, paraphrase, rewriting, extraction, Q&A, and chat are among the use cases that are supported.
Titan Text Lite
Very flexible and reasonably priced LLM. Appropriate in size for particular applications, perfect for text creation and optimisation jobs.
Maximum tokens: 4,000
Languages Spoken: English
Adjustment support: Indeed.
Copywriting and summarization are supported use cases.
Titan Text Embeddings
LLM that converts words into numerical formats.
Maximum tokens: 8,000
More than 25 languages are spoken.
No fine-tuning is supported.
1,536 embeddings
Semantic similarity, grouping, and text retrieval are supported use cases.
Titan Text Embeddings V2
Reduced storage and latency can be achieved by optimising LLM for high accuracy and retrieval performance at smaller dimensions.
Maximum tokens: 8,000
In pre-training, more than 100 languages
Supported fine tuning: None
Supported normalisation: yes
Embeds: 1024, 512, and 256
Supported use cases include using semantic similarity search to locate documents (e.g., for plagiarism detection), classifying labels into learnt representations based on data (e.g., for genre classification of movies), and enhancing the relevancy and quality of search results that are generated or retrieved.
Titan Multimodal Embeddings
Enables precise multimodal search and suggestion encounters.
Maximum tokens: 128
Maximum picture size: 25 MB
Languages Spoken: English
Adjustment support: Indeed.
Embeddings: 384, 256, 1,024 (default),
Search, recommendation, and personalisation are supported use cases.
Amazon Titan Image Generator
Using text prompts, create realistic, studio-quality photographs.
Maximum tokens: 77
Maximum file size for input: 25 MB
Languages Spoken: English
Adjustment support: Indeed.
Image manipulation, text to image generation, and picture variants are supported use cases.
You can now select the output vector size (256, 512, or 1024) with Amazon Titan Text Embeddings V2. Greater vector sizes result in more thorough answers, but they also take longer to process. Although they are less detailed, shorter vector lengths will speed up reaction times. Your storage expenses and the time it takes to search and retrieve document extracts from a vector database can both be decreased by using smaller vectors.
After evaluating the quality of the vectors produced by Amazon Titan Text Embeddings V2, we found that the accuracy given by vectors with 512 dimensions is roughly 99 percent of that of vectors with 1024 dimensions. 256-dimensional vectors maintain 97% of the precision. This implies that you can maintain roughly 97% of the precision offered by larger vectors while saving 75% on vector storage (from 1024 down to 256 dimensions).
Additionally, a better unit vector normalisation is suggested by Amazon Titan Text Embeddings V2 to help increase the precision of vector similarity measurements. Depending on your use case, you can select between normalised and unnormalized versions of the embeddings (normalised is more accurate for RAG usage situations). The process of scaling a vector to have a unit length or magnitude of one is known as normalisation. In order to prevent some vectors from dominating others because of their greater magnitudes, it is helpful to make sure that all of the vectors have the same scale and contribute equally throughout vector operations.
A multitude of application cases are well-suited for this new text embeddings approach. It can assist you in finding plagiarism in papers by performing semantic searches on them. For example, it may classify labels into learnt representations based on data and use that information to group movies into genres. Additionally, it can enhance the calibre and pertinence of obtained or produced search results by, for instance, employing RAG to suggest content according to user interest.
How embeddings contribute to increased RAG accuracy
Consider yourself an exceptionally skilled research assistant working on a large language model (LLM). Similar to brainiacs who can compose in a variety of imaginative language formats, LLMs derive their expertise from the enormous datasets they were educated on. This training set may be a little out of date or may not include all the information you require.
Here’s where RAG becomes useful. RAG functions as your personal assistant, retrieving pertinent data from a customised source, such as an organization’s knowledge store. RAG gives the LLM the most recent data so that it can produce the best response when the LLM is asked a question.
RAG employs embeddings to discover the most recent information available. Consider these embeddings, also known as vectors, to be extremely condensed summaries that encapsulate a text’s main idea. These summaries can be produced precisely using a top-notch embeddings model, like Amazon Titan Text Embeddings V2, which functions as a fantastic assistant by rapidly understanding the key elements of any document. This guarantees that RAG obtains the most pertinent data for the LLM, resulting in responses that are more precise and pertinent.
Consider it analogous to library research. The book’s pages are all indexed and represented by vectors. You could wind up with a stack of books that aren’t quite what you need if your search engine isn’t very good. However, if you have an excellent search system that comprehends the content (such as a high-quality embeddings model), you will find exactly what you’re looking for, which will greatly simplify the task of the LLM in coming up with the solution.
Overview of Amazon Titan Text Embeddings V2
For less storage and latency, Amazon Titan Text Embeddings V2 is optimised for good accuracy and retrieval performance at lower dimensions. Based on our measurements, vectors with 512 dimensions retain 99 percent of the precision offered by vectors with 1024 dimensions. A 256-dimensional set provides 97% of the accuracy.
Using Amazon Titan Text Embeddings V2
It is quite probable that you will engage in indirect interactions with Amazon Titan Text Embeddings V2 via Knowledge Bases for Amazon Bedrock. A RAG-based application is created with the hard lifting taken care of by Knowledge Bases. On the other hand, you can also call the model straight from your code using the Amazon Bedrock Runtime API.
In the US West (Oregon) and US East (North Virginia) AWS Regions, Amazon Bedrock is now offering the new Amazon Titan Text Embeddings V2 model.
