Coveo's Relevance Generative Answering Turns Search into a Conversation
Relevance Generative Answering (RGA) is Coveo’s latest AI offering that allows your users to ask complex search questions and receive generated responses using OpenAI’s GPT [3]. Coveo combines your content in a two-stage process [2] that provides context to the model, improving results. That process enables the model to respond with business-specific answers using industry-standard Large Language Model (LLM) technology.
How Do I Get Started with RGA?
The first stage in the process is to enable RGA to select the most relevant content using your query pipeline, which has been tuned with synonyms, conditions, and permissions. The second stage uses Coveo’s Semantic Encoder (Embedding) model, which retrieves related content that helps enrich and support the answer through more sources. The answer is validated in this stage, ensuring its quality meets or exceeds your standards.
This two-stage process helps users confidently use your Coveo solution's results. The user receives the generated answer, annotated with reference links to the original content, along with standard search results.
Watch this video for a high-level explanation of the Coveo AI Platform and RGA.
What Safeguards are in Place for My Data?
Coveo trains its LLM models without using your data, ensuring it can’t leak into the results of other users. Your data is maintained within a secure system that only sends what’s required to OpenAI for query and generative processing.
Your content is sent using secure network connections that are inaccessible to others. The transmission process includes a Semantic Encoder, which is a special type of index for your data, that’s located within your Coveo account. To protect your results, they’re filtered through your configured permission structure, which ensures only authenticated users can gain access, and only to what they need to know. You can read more about Coveo’s Relevance Generative Answering (RGA) data security policy here [0].
How Can RGA Help My Business?
By using Coveo’s RGA search instead of a standard search bar, your organization can provide users with a natural language interface that returns relevant, secure answers to their questions. Here are two specific use cases to give you a sense of its power.
Support Pages Powered by RGA
Help users find relevant solutions to issues that are partially documented across multiple files by applying RGA to support pages and message boards. RGA combines data from numerous sources to generate a cohesive, comprehensive explanation and provides links to the sources that informed the answer. This allows users to access more information, providing clarification when needed. The takeaway: users can get the info they need in less time, reducing frustration and reliance on call centers agents.
In-Product Assistants Powered by RGA
Provide users with the answers they need to stay productive, while avoiding toggling between browsers and search engines, by embedding RGA into a product panel. Indexing product documentation and other helpful information allows users to ask questions and receive answers specific to the product features they’re trying to use and understand, enhancing the customer experience.
Is Coveo’s RGA a Turnkey Solution?
Yes! Setting up and configuring the Relevance Generative Answering is straightforward. Begin by browsing the “models” section of Coveo’s admin interface. Then create both a Relevance Generative Answering and Semantic Encoder model. They aren’t intended to work separately as both must work together to provide a true generative experience. Each of these models require you, or the developer, to specify the sources to be used for ingestion. Filters that allow only specific content or ignore irrelevant source data can also be set up.
The Semantic encoder embeds your content, allowing it to search topics and find other related content. The time needed to conduct a search depends on the volume of data consumed. However, the models themselves handle the challenging work of training.
The query pipeline is connected to a search interface using Coveo’s new Atomic component architecture [4]. It includes the RGA component, which displays the users’ search responses. Coveo connects the search API responses to the component, enabling it to generate results.
What are RGA Best Practices?
When selecting content, make sure you’re providing relevant information. Consider what will inform the best answers when selecting what the model will ingest. Structure content into single-column paragraphs produces the best results. Content in table format can also be ingested, but it can be more challenging to process. Content can also be extracted using optical character recognition (OCR) from images.
Data requirements for the Semantic Encoder include:
- Indexing content in Coveo before creating the model
- Configuring Push API indexes with unique Permanent Ids
- Using English or translating content into English
- Using HTML or PDF files (videos are unsupported)
Enhance Your Internal- or Customer-Facing Websites and Products with Relevance Generative Answering
To get started on using Relevance Generative Answering in your business, reach out and we’ll help design a solution that enables your employees, suppliers, and customers to quickly get the information they need to take action.
References
[4] https://docs.coveo.com/en/atomic/latest/