Knowledge Retrieval

​

Key Features

• Vector Embeddings - Semantic search over document content
• RAG Integration - Automatically augment prompts with relevant context
• Multiple Sources - Upload documents, web pages, structured data
• Reranking - LLM-based result ranking for quality
• Citation Tracking - Know which documents contributed to answers

​

How to Enable

​

For Agents

1. Agents > Select Agent > Settings > Capabilities
2. Attach Knowledge Retrieval
3. Link to a Knowledge Base you’ve created

​

For MCP Gateways

1. Gateways > Select Gateway > Capabilities
2. Attach Knowledge Retrieval
3. Link to a Knowledge Base

​

Creating a Knowledge Base

​

How It Works

1. Document Upload - Documents processed into chunks with metadata
2. Embeddings - Each chunk converted to vector representation
3. Storage - Vectors stored in the vector database
4. Query - User query converted to vector
5. Search - Top matching documents retrieved via cosine similarity
6. Reranking (optional) - LLM reranks results for relevance
7. Augmentation - Top results injected into agent prompt

​

Usage Examples

​

Simple Search

"What are the main features of our product?"

​

Complex Question

"How do I troubleshoot authentication errors in the API?"

​

Document Analysis

"Summarize key findings from our Q4 financial report"

​

Configuration

​

Search Settings

{
  "knowledge_retrieval": {
    "embedding_model": "text-embedding-3-small",
    "top_k": 5,
    "similarity_threshold": 0.7,
    "enable_reranking": true
  }
}

​

Document Types Supported

​

Embedding Models

​

Search Examples

​

Keyword vs Semantic

"password reset"
→ Finds documents containing exact words

"How do I change my login credentials?"
→ Finds documents about password changes, auth, account settings

​

Multi-document

{
  "results": [
    {
      "document": "User Guide.pdf",
      "chunk": "Page 42",
      "content": "...",
      "score": 0.95
    },
    {
      "document": "FAQ.md",
      "chunk": "Section 3",
      "content": "...",
      "score": 0.87
    }
  ]
}

​

Reranking

• Retrieves top 20 results
• LLM evaluates each for relevance
• Reorders by actual relevance
• More accurate but slower

{
  "knowledge_retrieval": {
    "enable_reranking": true
  }
}

​

Cost

​

Rate Limits

• 1000 searches per hour per knowledge base
• 500 document uploads per day per account
• 1M total embeddings per account

​

Best Practices

​

Document Preparation

• Clean PDFs work better than scanned images
• Use clear document structure
• Include metadata and dates
• Break large documents into sections

​

Search Queries

• Be specific and descriptive
• Use natural language
• Ask complete questions
• Avoid single-word searches

​

Knowledge Base Organization

• Group related documents
• Use consistent naming
• Update documents regularly
• Remove outdated content

​

Common Use Cases

​

Product Documentation

"How do I reset my password?"
→ Returns relevant FAQ section

​

Internal Knowledge

"What's our refund policy?"
→ Returns policy document

​

Research

"What are recent findings on AI safety?"
→ Returns relevant paper sections

​

Troubleshooting

​

Poor Search Results

• Verify documents uploaded correctly
• Try more specific queries
• Enable reranking for better results
• Check similarity threshold

​

Slow Search

• Reduce top_k value
• Disable reranking for speed
• Use smaller embedding model

​

Upload Failed

• Check file format is supported
• Verify file is valid (not corrupted)
• Try smaller file size
• Check account storage limit

​

Next Steps

• Create Knowledge Base
• Upload Documents
• Vector Search
• Reranking