Skip to main content

Knowledge Assist

Knowledge Assist enables CX Assistants to retrieve and use information from curated knowledge bases to answer questions. By connecting Assistants to knowledge bases, you can ensure responses are grounded in accurate, up-to-date information rather than relying solely on the LLM's training data.

Overview​

Knowledge Assist combines multiple search techniques to find the most relevant information:

  1. Text Search: Traditional keyword-based matching for exact terms
  2. Semantic Search: Vector-based similarity search to find conceptually related content
  3. Hybrid Ranking: Reciprocal Rank Fusion combines results from both methods for optimal relevance

How It Works​

1. Knowledge Base Setup​

Create a knowledge base with URLs to crawl (web pages, help articles, documentation) and/or files (PDFs, documents). The web crawler extracts content and stores it in a searchable index.

2. Connect Assistant to Knowledge Base​

Configure an Assistant with a knowledge_base_code. This automatically:

  • Adds the retrieve_knowledge tool to the Assistant
  • Injects instructions for using the knowledge base
  • Enables grounded responses based on retrieved content

3. Retrieval During Evaluation​

When the Assistant receives a question:

  1. The LLM calls the retrieve_knowledge tool with a search query
  2. The tool performs hybrid search (text + semantic) on the knowledge base
  3. Results are ranked using Reciprocal Rank Fusion
  4. Relevant documents are returned to the LLM as context
  5. The LLM generates a response based on the retrieved content

Key Features​

Combines text search (BM25) and semantic search (vector embeddings) for comprehensive retrieval:

  • Text search excels at finding exact matches and specific terminology
  • Semantic search finds conceptually similar content even with different wording

Context-Aware Retrieval​

The tool automatically:

  • Limits results to fit within the model's context window
  • Deduplicates overlapping content
  • Preserves document metadata (title, source URL) for citations

Web Crawling​

The built-in web crawler:

  • Extracts content from configured URLs
  • Processes linked pages (configurable depth)
  • Handles PDF files and other document types
  • Tracks crawl status and notifies on completion

Use Cases​

Customer Support​

Connect support Assistants to your help center articles:

  • Agents get accurate answers from official documentation
  • Responses include links to source articles
  • Knowledge stays current with automatic re-crawling

Internal Knowledge Base​

Ground Assistants in internal policies and procedures:

  • HR policy lookups
  • Technical documentation
  • Standard operating procedures

Product Information​

Connect sales or support Assistants to product documentation:

  • Feature explanations
  • Pricing information
  • Troubleshooting guides

Prerequisites​

Before setting up Knowledge Assist:

  1. Account Configuration: Knowledge Assist must be enabled for your account (deploy_knowledge_assist flag)
  2. Content Sources: Have URLs or files ready to add to the knowledge base
  3. Assistant Configuration: Create or identify the Assistant to connect

Next Steps​