Knowledge Copilot indexes the relevant enterprise knowledge sources and, using a proprietary engine, retrieves the items holding the information relevant to answering a given question and generates an answer while running through quality guardrails to ensure relevance and correctness.

The engine leverages Precision RAG, a Retrieval Augmented Generation approach of prompting LLMs (Large language models) with a given question and context from which to provide their answer. Precision RAG has guardrails in place to ensure the relevance and correctness of the provided answer.

A search engine retrieves a list of relevant links or items most related to a given query which, in most cases, needs to be short and specific. An answer engine provides an actual answer to a given question, which can be long and complex, while noting the items this answer is based on.

Unlike most search engines employed in RAG setups, which employ a singular, or sometimes a small combination of models for the retrieval part of the process, xFind employs a broad setup of models which complement each other.

This enables Knowledge Copilot to:

• Successfully deal with queries much longer than classic search
• Successfully retrieve from more complex sources, such as past cases
• Be much more focused in retrieving the relevant items at the top of the list in order to feed the answering LLM

Retrieval Augmented Generation – an approach of prompting LLMs (large language models) with a given question and context from which to provide an answer.

Precision RAG provides trusted, real-time answers to complex questions, removing the need to deduce these answers from the various existing sources in which they reside.

Precision RAG is a method for turning complex queries and disparate data sources into precise answers – the ideal technology for handling support case data and complex customer questions.

Many don’t realize the complexity and challenges involved in developing AI-based question-answering solutions. They often underestimate the intricacies of handling unstructured data and ensuring accurate and cost-effective retrieval and response generation.

The main challenges include:

• Data source complexity: Enterprises often rely on unstructured data from past cases, tickets, emails, etc., which requires advanced parsing pipelines for consistency.
• Search inaccuracy: Pre-trained embeddings from models like OpenAI, Claude, Mistral, and LLama are not precise enough for domain-specific data, leading to inaccuracies in retrieval.
• Cost: Reducing costs at scale involves focusing queries only on necessary information, requiring a highly precise search engine.
• Search failure detection: It’s crucial to know when the system doesn’t have an answer, especially in high-stakes support scenarios, to manage customer expectations and provide accurate information.

The effectiveness of Retrieval Augmented Generation (RAG) in enterprise settings depends primarily on the quality of the search engine. A powerful search engine ensures precise and relevant data retrieval, which is critical for generating accurate answers, regardless of the generative LLM’s capabilities.

Knowledge Copilot leverages advanced technology to manage these challenges at scale, including:

• Data preparation pipelines: These pipelines can handle any data format and type, making it ready for both search engines and LLMs.
• Hybrid search engine setup: xFind integrates various language model types and methods, allowing for a plug-and-play setup with minimal time and no training data required.
• Guardrails: Real-time assessment ensures queries are relevant to the knowledge domain and that sources contain accurate information, minimizing hallucinations and errors.

The proprietary guardrails in xFind’s technology automatically evaluate whether a query is relevant to the knowledge domain and if the sources can provide an accurate answer. This helps minimize errors related to topic mismatches or missing data, ensuring more reliable and accurate responses.

Knowledge Copilot implements guardrails that are able to detect:

• If a query is in the general domain of the customer’s knowledge, to avoid answering irrelevant questions such as ‘how do I walk my dog’
• If the results retrieved include information that could be relevant to answering the question
• Focus the items retrieved to a very specific set, to help the LLM with finding the relevant information to use when building the response

Any textual knowledge source, including voice transcriptions. Out-of-the-box sources include:

• Salesforce
• Zendesk
• Mindtouch
• KnowledgeOwl
• StackOverflow
• Salesforce Knowledge
• Zendesk articles
• Atlassian (Confluence, Jira)
• Notion
• Paligo
• GetGuru
• Gitbook
• Magentrix
• HigherLogic
• Doc360
• Slack
• Sitecore
• MS Teams

New customers typically onboard and start seeing AI insights from core implementation within 45 days.

SupportLogic implementation is a short and straightforward process after initial implementation.

Implementation is a flat fee. Contact us for a price estimate.

SupportLogic employs a usage-based pricing model. We charge an implementation fee for each module and then the actions you take in the system incur charges based on usage and value. This ensures you only pay for the parts of the platform that you use.

Contact us for a full estimate of annual pricing based on your initial case volume and selected use-cases.